Arxiv:1501.02273V2 [Cond-Mat.Mes-Hall] 16 Feb 2015

Exciton band structure of monolayer MoS2

Fengcheng Wu,1 Fanyao Qu,1, 2, ∗ and A. H. MacDonald1, † 1Department of Physics, University of Texas at Austin, Austin, TX 78712, USA 2Instituto de F´ısica, Universidade de Bras´ılia, Bras´ılia-DF70919-970, Brasil (Dated: February 17, 2015)

We address the properties of excitons in monolayer MoS2 from a theoretical point of view, showing that low-energy excitonic states occur both at the Brillouin zone center and at the Brillouin-zone corners, that binding energies at the Brillouin-zone center deviate strongly from the (n − 1/2)−2 pattern of the two-dimensional hydrogenic model, and that the valley-degenerate exciton doublet at the Brillouin-zone center splits at finite momentum into an upper mode with non-analytic linear dispersion and a lower mode with quadratic dispersion. Although monolayer MoS2 is a direct-gap semiconductor when classified by its quasiparticle band structure, it may well be an indirect gap material when classified by its excitation spectra.

PACS numbers: 71.35.-y, 78.67.-n, 73.22.-f

I. INTRODUCTION mode with non-analytic linear dispersion. This unusual pattern is due to valley coherence established by electron- In monolayer form the group VI transition-metal hole exchange interactions. dichalcogenides (TMDs) like MoS2 are an interesting Low energy exciton states appear both near the class of semiconductors, and one that has recently re- Brillouin-zone center and near the Brillouin-zone corners. ceived considerable attention.1–17 In these materials con- There are two distinct types of Brillouin-zone corner exci- duction and valence bands are both dominantly d- tons. One type has electrons and holes in opposite valleys electron in character and have band extrema located (K, K0), while the other has holes in the Γ valley and at the triangular lattice Brillouin-zone-corners K and electrons in either the K or K0 valley. Although mono- K0. Because their structure breaks inversion symme- layer MoS2 is a direct-gap semiconductor as judged by its try, coupling is allowed between real spin and valley band structure, because of these Brillouin-zone corner ex- pseudospin3 and gives rise to valley-dependent optical citon states, we propose that it may well be an indirect selection rules.4–6 Because of relatively large carrier effec- gap material when judged by its excitation spectra. tive masses, reduced screening, and carrier confinement Our paper is organized as follows. In Section II we in a single atomic layer, their electron-hole interactions describe the model we employ for quasiparticle bands are much stronger than in conventional semiconductors. and for electron-hole interactions, and in Section III we Monolayer TMDs therefore host exceptionally strongly present our results. We conclude in Section IV with a bound excitons and trions that have been extensively summary and brief discussion. studied both experimentally and theoretically.7–15 In this article we report on a theoretical study of exciton energies and wave functions in MoS2 as a func- II. THEORETICAL FORMULATION tion of momentum across the full Brillouin zone. We identify important aspects of 2D-TMD exciton physics Exciton states are obtained by solving a two-particle that are controlled by mirror, three-fold rotational, and problem with attractive interactions between one conduc- time-reversal discrete symmetries. We calculate the opti- tion band electron and one valence band hole. Because cal conductivity, which reflects the properties of excitons the valence and conduction band edges are dominated with zero center-of-mass momentum and exhibits a set of by Mo atom d-orbitals we use a five band d-orbital tight- peaks split by electron-hole binding energies as usual, but binding model, detailed in Appendix A and illustrated in also by large valence band spin-orbit coupling energies. Fig. 1, for the quasiparticle bands of the TMD semicon- The spin splitting of valence band is conventionally used ductor ground state. Provided that the typical separa- to classify absorption peaks into A and B series. The tion between the electrons and holes in exciton states is arXiv:1501.02273v2 [cond-mat.mes-hall] 16 Feb 2015 exciton energy pattern is distinctly different from that of at least several lattice constants we can assume that the a 2D hydrogenic model. In particular the four A2p states electron-hole interaction strengths are dependent mainly are lower in energy than the corresponding 2s states, and on the separation between atomic sites and not on the d- not degenerate. orbital character on that site. These considerations lead Finite-momentum excitons are optically inactive, but to a Hamiltonian of the form can nevertheless play an important role in hot carrier 18–23 relaxation and in valley dynamics. We find that for H = H0 + HI, both A and B excitons, the valley degenerate states at 1 X † † (1) the Brilliouin-zone center split at small momentum into H = V 0 a a a 0 0 a . I 2 |R~−R~ | Rν~ R~ 0ν0 R~ ν Rν~ a lower mode with quadratic dispersion and an upper R,~ R~ 0 2

5 Exciton states with center of mass momentum Q~ can be expanded in terms of one-electron/one-hole states: 4 X ~ ~ ~ |χiQ~ = ψQ~ (v, c, k) |v, c, k, Qi, (3) 3 v,c,~k L

eV 2 † H ~ ~ c2 ¢ c2 where |v, c, k, Qi = b b~ |Gi, |Gi is the neutral ¶ K (~k+Q~ )c kv c1 M c1 1 semiconductor ground state, and the sums are over all G K † valance (v) and conduction (c) bands. b~kn and b~ are v v kn 0 1 1 quasiparticle operators for band n at momentum ~k. The v v 2 2 wave vector ~k + Q~ is understood to be reduced to the -1 K G M K¢ Brillouin-zone. The exciton center-of-mass momentum Q~ is also understood to be conﬁned to the Brillouin-zone FIG. 1: (Color online) Quasiparticle band structure of mono- and is a good quantum number. Like the quasiparticles, layer MoS2. The solid curves were obtained using the excitons have a band structure. To characterize an ex- Quantum ESPRESSO package28 with fully relativistic pseu- citon state, we deﬁne its ~k-space probability distribution dopotentials under the Perdew-Burke-Ernzerhof generalized- function as gradient approximation, and a 16×16×1 ~k-grid. The dashed ~ X ~ 2 curves were calculated from the tight-binding model, with PQ~ (k) = |ψQ~ (v, c, k)| . (4) cyan (red) representing states that are even (odd) under mir- v,c ror operation with respect to the Mo plane. v and c la- 1,2 1,2 The eigenvalue problem for the Hamiltonian matrix bel the bands close to the valence and conduction band edges near the K and K0 points. The inset shows the hexagonal projected onto this subspace is a Bethe-Salpeter (BS) Brillouin zone (pink) associated with the triangular Bravais equation. Its solution determines the exciton energies lattice of MoS2 and an alternate rhombohedral primitive zone EQ~ and wave functions. The Hamiltonian matrix (black), and labels the principle high-symmetry points in reciprocal space. Note that the valence band maxima at Γ is hv, c,~k, Q~ |H|v0, c0,~k0, Q~ i only slightly lower in energy than the valence band maxima 0 cc 0 0 0 0 ~ ~ ~ at K,K =δvv δcc δ~k~k0 ε(~k+Q~ )c − ε~kv − D − X vv0 (k, k , Q), (5)

where ε~kn denotes the quasiparticle energy. We view the where H0 is the Hamiltonian for independent d electrons two-dimensional bands predicted by density-functional and H describes their interactions. In Eq. (1) a† (a ) theory electronic structure calculations, illustrated in I Rν~ Rν~ is the electron creation (annihilation) operator for orbital Fig. 1 as solutions of the neutral semiconductor single- ν at Mo site R~ and ν = (o, s) includes both orbital o and particle Dyson equation including all many-body self- spin s labels. We follow recent work24,25 by using an energy effects except for finite lifetimes. It follows that in interaction potential of the Keldysh form,26,27 the exciton calculation we need to account only for cor- rections due to electron-hole interactions. It will, how- πe2 ever, be necessary to correct for the well-known ten- VR = [H0(R/r0) − Y0(R/r0)], (2) dency of density-functional-theory bands to underesti- 2r0 mate semiconductor gaps. As discussed later, this con- to account for the finite width of the TMD layer and sideration motivates shifting the calculated excitation en- the spatial inhomogeneity of the dielectric screening energy spectrum rigidly to match experimental optical ab- vironment. This interaction gives a good description of sorption spectra. the nonhydrogenic Rydberg series observed in monolayer In Eq. (5), D and X are respectively the direct and 24,25 WS2. In Eq. (2) is an environment-dependent di- exchange two-particle matrix elements: electric constant, r0 is a characteristic length related to cc0 0 1 † † the width of a single TMD layer, and H0 and Y0 are re- 0 ~ ~ ~ Dvv (k, k , Q) = V~k−~k0 U~ ~ U~k0+Q~ 0 U~ 0 U~k 0 , spectively Struve and Bessel functions of the second kind. N k+Q cc k v v Unless otherwise stated, we chose = 2.5, which corre- cc0 0 1 † † X 0 (~k,~k , Q~ ) = V U U U U , vv Q~ ~k+Q~ ~k cv ~k0 ~k0+Q~ v0c0 sponds to MoS2 lying on a SiO2 substrate and exposed N to air. r depends on and we took r = 33.875A˚/ (6) 0 0 where U is the unitary matrix which diagonalizes the from Ref. [10] . The onsite interaction is regularized by ~k setting V = UV with a equal to the lattice param- quasiparticle Hamiltonian (see Appendix A), N is the 0 R=a0 0 number of unit cells in the finite system over which we eter of MoS2, and the parameter U is taken to be 1 for P i~q·R~ results presented below. The dependence of our results apply periodic boundary conditions, and V~q = e VR on the value chosen for the dimensionless parameter U, is the lattice Fourier transform of the interaction poten- which accounts for screening of on-site potentials by re- tial. Note that V (~q) = V (~q + G~ ) for any reciprocal lat- mote bands, will be discussed later. tice vector G~ . For the special case Q~ = 0, the exchange 3

                      (a)               (b) 0.30                   A 2.3                                                                                                                             0.25               2.2                                              ) )    A2s          0.20     eV eV    K  ( ( A2p{   K K '    b B E

 E  K '    4 B 2 13, K  K K  0.15     K       3  K A A  K  2 0.10 2s  K 1.9 1 K K ' K 27 33 39 45 51 57  M K N1/2

FIG. 2: (Color online) (a) Energies of Type III excitons as a function of center-of-mass momentum Q~ . This ﬁgure is based on a calculation performed using a 45 × 45 ~k-grid. The lines were added as a guide to the eye. Solid (dashed) lines represent states that are doubly(singly) degenerate. The labels of the excitons with Q~ = 0 are explained in the main text. Excitons ~ 1 2 ~ ~ with Q = K are labeled by χK , χK and so on in ascending order of energy. The left inset is a k-space map plot of PQ~ (k) (see ~ 2 Eq. (4)) for the Q = 45 M exciton in the lower-energy branch evolving from A. The right inset schematically illustrates the 1 2 3 dominant electron-hole transitions which contribute to the χK , χK and χK exciton states. (b) Binding energy Eb for A, B and 1/2 A2s excitons at Q~ = 0 as a function of N , where N is number of ~k points. term X vanishes because of the orthogonality property boundary conditions that restrict ~k to a regular discrete † grid in the primitive zone illustrated in Fig. 1. Our main U~ U~k cv = 0. We remark that exchange term survives k results are summarized in Fig. 2(a), which shows the en- even at Q~ = 0 in models in which electron-hole interac- ergies of Type III excitons as a function of center-of-mass tions depend not only on electron-hole separation, but ~ also on orbital character22. momentum Q. To test the convergence of our calculations with respect to the ~k-space sampling density, we plot the binding energies of low-energy excitons as a func- III. EXCITON BAND STRUCTURE tion of periodic system size in Fig. 2(b). We start by an- alyzing excitons at and close to the Γ point (Q~ = 0), and then discuss the nearly degenerate low-energy excitons Monolayer MoS has mirror symmetry29 with respect 2 ~ 0 to the Mo plane. Quasiparticle band spinors can be clas- Q = K,K . ~ ~ siﬁed by this symmetry M|k, ni = −imn|k, ni, where the mirror number mn = +(−) for mirror even (odd) bands as shown in Fig. 1. Using mirror numbers, we can group exciton states into three decoupled types: (1) A MoS2 has a 3-fold rotational symmetry which can ~ ˆ Type I exciton is formed by promoting an electron from be used to classify excitons with Q = 0: C3|χi = exp(−i 2π L)|χi, where the quantum number L takes on a mirror-even valance band (mv = +) to mirror-odd con- 3 the discrete values L = −1, 0, 1. An exciton state |χ+i duction bands (mc = −); (2) Type II is similar to Type I but with (m , m ) = (−, +); (3) For a Type III exciton, with L = 1 has a time reversal (TR) partner |χ−i with v c ˆ mv = mc = ±. Only Type III excitons can be optically the opposite L = −1. The combination of C3 and TR bright. Exchange terms vanish in Type I and II excitons symmetries guarantees that the TR pair |χ+i and |χ−i because their valence and conduction bands have oppo- are degenerate in energy. Breaking either Cˆ3 or TR site mirror numbers. For a Type III exciton, the two symmetry can lift this degeneracy30–35. The optical se- sectors mv = mc = + and mv = mc = − are coupled lection rule for circularly polarized light is related to 4,36 by exchange terms, but not by direct terms. In the fol- Cˆ3 symmetry : hχ+|ˆj−|Gi = hχ−|ˆj+|Gi = 0, where lowing, we restrict our discussion to Type III excitons, ˆj± = ˆjx ± iˆjy is the current operator. It follows that although many of the points we make apply equally well polarization-dependent optical studies can can be used to to Type I and Type II excitons. infer the L quantum number of a bright exciton. L = 0 We have solved the BS equation by applying periodic excitons are optically inactive hχ0|ˆj±|Gi = 0. 4

with the estimate in Ref. [10]. As shown in Fig. 2(b), 25 (a) B A A 2p a 33 × 33 ~k-grid already provides good convergence for 20 the lowest energy A and B excitons whereas, because

) more weakly bound excitons have sharper structure in /h 2

e 15 momentum space as illustrated in Fig. 3(e), the A2s ex- (

xx B2s citon requires a finer ~k-grid for convergence. σ 10 The energies of s-wave excitons in the A(B) series de- Re −2 A2s viate strongly from the (n − 1/2) pattern of 2D hydro- A3s 5 genic models. This is partly due to the effective electron- hole interaction potential(Eq. (2)), which differs from the 0 standard Coulomb interaction because of the finite width 1.8 1.9 2.0 2.1 2.2 2.3 of the TMD layer. We also find that A states have a ħω(eV) 2p lower energy than A2s states. This anomalous energy ordering is consistent with recent experimental and theoret- (b) (c) (d) (e) 37 ical studies of monolayer WS2 . More interestingly, the K ' A2p states do not have the 4-fold degeneracy expected in two-valley systems. In fact, there are two non-degenerate A states within each valley, as illustrated in Fig. 3(c) K 2p  and 3(d). This feature results from the dependence of band state wave functions on momentum direction near K and K0 valleys and is closely related to similar properties of excitons in massive Dirac equation band models, FIG. 3: (Color online) (a) Real part of the optical conductiv- which we explain in detail in Appendix B. The A2p states ity with (solid red curve) and without (dashed green curve) are not optically bright in one-photon spectra, but can be electron-hole interactions. For these calculations the BS equa- detected using two-photon techniques like those achieved tion was solved on a 51×51 ~k-grid and transitions were broad- in recent experiments.37,38 We therefore expect that the ened by 20meV. The solid green arrow indicates the quasipar- energy splitting within the A2p states can be experimen- ticle band gap. The two dashed gray arrows mark the ener- tally measured. gies of the A2p excitons. Note that we have rigidly shift the excitation energy spectrum by a constant, so that the A exciton energy is at 1.93eV as measured by photoluminescence B. Valley Coherence 6,9–11 ~ ~ experiments . (b)-(e) k-space maps of PQ~ =0(k) for low- energy excitons. (b)The L = 1 exciton |A+i. (c)-(d) Two Valley coherence can be externally generated using lin- non-degenerate A2p excitons in valley K. (e) The L = 1 A2s 39 exciton. early polarized light , and can also be intrinsically induced by electron-hole exchange interactions. By treat- ing finite Q~ terms in the BS equation(Eq. (5)) as a first- A. Optical Conductivity order perturbation acting on valley-degenerate excitons, we arrive at the effective Hamiltonian 2 2 Exciton states at Q~ = 0 are most easily studied exper- eff ~ Q H ~ = ~ω0 + ] τ0 + J ~ τ0 imentally because they contribute to the optical conduc- Q 2M Q (7) tivity. Fig. 3(a) plots the real part of the in-plane optical +JQ~ [cos(2φQ~ )τx + sin(2φQ~ )τy]. conductivity which has a number of clear features. Peak ~ A stems from the doubly-degenerate excitons expected Here ω0 is the exciton energy at Q = 0, and M is the ex- from the symmetry analysis given above. Fig. 3(b) illus- citon mass. τ0 and τx,y are respectively identity and off- diagonal Pauli matrices in valley space. J τ originates trates P (~k) (see Eq. (4)) for |A i, the L = 1 exciton Q~ 0 Q~ =0 + from intra-valley exchange interactions, while inter-valley of peak A. |A i is dominated by electron-hole transitions + exchange interactions act as an in-plane pseudo-magnetic from valence band v to conduction band c in valley K 1 2 field in the valley space and are captured by the second (see Fig. 1), while its TR partner |A i is primarily com- − line of Eq. (7). The dependence of Heff on φ , the orien- posed of similar transitions in the opposite valley K0. Q~ Q~ Excitons in the B series are similar to those in A, and tation angle of the 2D vector Q~ , follows from the wave- are dominated by transitions from band v2 to c1 in valley vector dependence of the conduction and valence band K (K0). The lowest energy A and B excitons are analo- states near K and K0. Ref. [22] and [30] have studied gous to the 1s states of a 2D hydrogenic model. Fig. 3(a) the inter-valley exchange interaction and show that, also shows 2s and 3s peaks identified in the A series, and J ∝ |ψ (0)|2Q2V , (8) a 2s peak identified in the B series. Q~ eh Q~ 2 Our calculation predicts that the lowest energy A ex- where |ψeh(0)| is the probability that an electron and citons have a binding energy ∼ 0.3 eV, in agreement a hole overlap spatially. In the small Q limit, VQ~ ∝ 5

1/(Q(1 + r0Q)) for the potential in Eq. (2). Therefore, 5 JQ~ scales linearly with Q in the long wave-length limit. We note that while inter-valley exchange interaction en- 0 dows ﬁnite-momentum excitons with chirality I = 2 as pointed out by Ref. [30], intra-valley exchange22,35,40 is -5 also important especially in regard to the exciton energy )-E(A)( meV) dispersion. 3 K -10

Equation (7) is derived from the massive Dirac model E(χ approximation to the quasiparticle band structure near -15 K and K’.22,30 Our lattice calculation verifies this low- energy effective theory. Fig. 2(a) shows that there are 1.0 1.5 2.0 2.5 3.0 two non-degenerate energy branches which evolve from ϵ the double-degenerate Q~ = 0 A, B and A2s excitons. In each energy branch, the exciton state is a coherent super- FIG. 4: The energy difference between the large-momentum 3 position of direct excitons at the two valleys, as demon- exciton χK and zero-momentum exciton A as a function of strated in the left inset of Fig. 2(a). The lower and upper dielectric constant . The calculation was performed on a ~ 3 energy branch have respectively quadratic and linear dis- 45 × 45 k-grid. χK is the triplet exciton state with a hole in persion in the long wave-length limit, in agreement with Γ valley and an electron in K valley. the prediction of Eq. (7). Unlike their s-wave cousins, branches evolving from valley-degenerate Q~ = 0 A2p excitons remain doubly degenerate at finite momentum be- 2 cause |ψeh(0)| is zero for p-wave excitons and exchange interactions therefore vanish. Photons with the energy of an A or B exciton can at valley, the corresponding excitons can be classified as sin- most provide a momentum with magnitude ∼ ωA(B)/c, where c is the speed of light. According to our calcu- glets and triplets according to their spin configurations. lation, this momentum corresponds to an energy split- Singlets experience strong electron-hole exchange inter- ting of 0.4(0.5) meV between the two energy branches actions, because the valence band maximum at Γ and evolving from A(B), and a period of 11(8)ps for Rabi the conduction band minimum at K have the same dz2 oscillation between the two valleys. We conclude that orbital character. The exchange energy is proportional interaction-induced valley coherence provides an impor- to VQ~ =K = (U − 1.28)VR=a0 . The exchange interactions tant mechanism for valley depolarization21–23, in addi- effectively vanishes for triplets because of their particular tion to that provided by impurity scattering. spin structure. In our calculation, U is set to be 1 and the exchange energy for singlets is therefore negative. There- 1 3 fore, we can identify χK as the singlet state and χK as one of the triplet states. Depending on the value of U C. Brillouin-zone corner Excitions used in our model, the lowest energy excitons can occur at Q~ = 0, corresponding to a direct gap system, or at Large-momentum excitons composed of electrons and Q~ = K,K0 corresponding to an indirect gap system. Be- holes in opposite valleys can have an energy similar to cause the appropriate value which should be used for U those with zero momentum. In Fig. 2(a), the χ2 and χ4 K K depends on electronic correlations at the atomic level and ~ excitons have center-of-mass momentum Q = K when on screening by remote bands not included in our calcu- reduced to the first Brillouin-zone. They are dominated lation, we are not able to reach a definitive conclusion respectively by transitions between valence band holes of 1 as to whether or not the singlet χK is the lowest energy v1 and v2 states in valley K and conduction band elec- exciton. However, the energy of the triplet χ3 does not 0 K trons of c1 and c2 states in the opposite valley K . The suffer from such uncertainty. The valence band effective 2 4 ~ 42 χK and χK excitons are the Q = K counterparts of the mass at the Γ point is heavier than that at the K point . A and B excitons and have nearly the same energy. The Electron-hole pairs are therefore bound more strongly in small differences have two origins, the energy splitting be- χ3 than in the A exciton, which compensates for the 41 K tween c1 and c2 bands , and a change in the exchange energy difference (68meV) between the topmost valence interaction. bands at Γ and K points, and makes the energies of the 3 However another set of excitations appear at the same triplet χK and A excitons very close to each other. By crystal momentum. Low energy excitations at Q~ = K decreasing the dielectric constant to 1, we find that the 3 also originate from holes in the Γ valley and electrons in triplet χK becomes lower in energy than the A exciton by the K valley, as illustrated in the right inset of Fig. 2(a). 19meV, as illustrated in Fig. 4. The energetic ordering 1 3 In our lattice calculations χK and χK are such excitons. of direct and indirect exciton states in single-layer TMDs If we neglect the spin splitting of both topmost valence could therefore depend on the two-dimensional system’s bands at Γ and the lowest conduction bands at the K three-dimensional dielectric environment. 6

even ~ IV. SUMMARY AND CONCLUSIONS For H0 (k), we adopt a model constructed in Ref. 43. The construction uses point-group symmetries to mini- In summary, we have constructed a lattice model based mize the number of parameters, and the parameters are on Mo d-orbitals to study exciton band sturcture of fitted from first-principle energy bands. An explicit form even ~ monolayer MoS2. Zero-momentum excitons have non- of H0 (k) with hoppings up to third-nearest-neighbor hydrogenic energy series, because screening effect has a is given in Eqs. (13) to (24) of Ref. 43. We generalize the spatial dependence, and band edges in K and K0 val- symmetry-based method used in Ref. 43 to construct the odd ~ ley are described by massive Dirac equation. The energy remaining part of the Hamiltonian H0 (k), and its form splitting within A2p excitons remains to be measured us- is: ing methods such as two-photon technique. The exciton ! h (~k) h (~k) band structure exhibits non-degenerate energy branches Hodd(~k) = x xy , (A3) 0 ∗ ~ ~ evolving from valley-degenerate bright excitons, indicat- hxy(k) hy(k) ing valley coherence. Such low-energy branches are well separated from the continuum spectrum, which justifies in which the application of low-energy effective theory22,30. We h (~k) =O + 2t cos 2α + (t + 3t0) cos α cos β find that low-energy excitons can possess a large momen- x 1 0 tum, either with electron and hole in opposite valleys (K, +4s cos 3α cos β + (3s − s) cos 2β K0), or with hole in Γ valley and electron in K(K0) valley. +2u cos 4α + (u + 3u0) cos 2α cos 2β, Large-momentum low-energy exciton states can provide h (~k) =O + 2t0 cos 2α + (t0 + 3t) cos α cos β relaxation channels for bright excitons, and reduce pho- y 1 0 0 toluminescence quantum efficiency. +4s cos 3α cos β + (3s − s ) cos 2β +2u0 cos 4α + (u0 + 3u) cos 2α cos 2β, ~ hxy(k) =4itxy sin α(cos α − cos β) (A4) V. ACKNOWLEDGMENTS √ + 3(t0 − t) sin α sin β √ This work was supported by the DOE Division of Ma- +2 3(s0 − s) sin α sin β(1 + 2 cos 2α) terials Sciences and Engineering under grant DE-FG03- +4iuxy sin 2α(cos 2α − cos 2β) 02ER45958, and by the Welch foundation under grant √ TBF1473. We thank Texas Advanced Computing Cen- + 3(u0 − u) sin 2α sin 2β, ter(TACC) for providing computer time allocations. √ 1 3 (α, β) =( k a , k a ). 2 x 0 2 y 0 Appendix A: Explicit form of tight-binding model Hoppings in real space up to third nearest neighbors are included. The numerical value of the parameters in unit We approximate the quasiparticle-Hamiltonian matrix of eV is: H~ by a tight-binding model which generalizes Ref. [43] k O1 =3.558, from three to five d-bands: 0 t = − 0.189, t = −0.117, txy = 0.024, ~ ~ ~ 0 (A5) H~k = λL · S + I2 ⊗ H0(k). (A1) s = − 0.041, s = 0.003, 0 u =0.165, u = −0.122, uxy = −0.140, The first term λL~ · S~ describes the on-site atomic spin- orbit coupling of Mo d orbitals, where L~ and S~ are re- which are obtained by fitting to first-principle calcula- spectively the orbital and spin angular momentum. The tions as shown in Fig. 1. coupling constant λ = 0.073eV, which was adjusted to fit H~k is diagonalized by the unitary matrix U~k (Eq. (6)), the valence-band spin splitting at the K point as detailed and the corresponding eigenvalues are quasiparticle en- ~ ergies ε . in Ref. [43]. The second term I2 ⊗ H0(k) is spin indepen- ~kn dent, where I2 is a 2 × 2 identity matrix in spin space, and H (~k) is a 5 × 5 matrix in orbital space. Because 0 Appendix B: Massive Dirac model for excitons of the mirror symmetry with respect to the Mo plane, H (~k) is block-diagonal: 0 In this appendix we study excitons in the massive Dirac Heven(~k) 0 model, and show that this simple model captures many H (~k) = 0 . (A2) ~ 0 odd ~ important features of the Q = 0 excitons in monolayer 0 H0 (k) 0 MoS2. In the vicinity of K or K point, the k · p Hamil- even ~ tonian for two bands (c2, v1) is described by the massive H0 (k) is a 3 × 3 matrix in the bases Dirac model3: odd ~ {|dz2 i, |dxyi, |dx2−y2 i}. Similarly, H0 (k) is a 2 × 2 ~ matrix in the bases {|dxzi, |dyzi}. Hτ (k) = ~vF k[cos(Φ~k)σx + sin(Φ~k)σy] + ∆σz, (B1) 7

Eb(eV) @ ϵ=2.5 MD LM 1s(l = 0) 0.345 0.301 ● ● 4 ● ● 0.6 ● ● ● 2p(l = -1) 0.159 0.150 ● ● ● ● ● ● 2p(l = +1) 0.143 0.125 2s(l = 0) 0.118 0.099 3

0.4 * ■

eV /Ry b

b E E 2 ■ ● 1s(l=0) ■ 2p(l=-1) ■ 0.2 ◆ 2p(l=+1) ■ 1 ▲ 2s(l=0) ■ ■ 4 ■ ▲ ▲ ◆■▲ ◆▲■ ◆▲■ ◆■▲ ◆■▲ ◆■▲ ◆▲ ◆▲ ◆▲ ◆▲ ◆▲ ◆ ◆ 9 0 0 1.0 1.2 1.4 1.6 1.8 2.0 2.2 2.4 0.0 0.2 0.4 0.6 0.8 1.0 1.2 α

FIG. 5: (Color online) Binding energy Eb in massive Dirac FIG. 6: (Color online) Binding energy Eb in massive Dirac model for 1s, 2s, and 2p excitons in valley K (τ = 1) as a model with standard Coulomb interaction(r0 = 0,F (q) = 1) function of dielectric constant . The parameter values are for 1s, 2s, and 2p excitons in valley K (τ = 1) as a function of ﬁne structure constant α. ~vF = 1.105eV × 3.193A,˚ ∆ = 0.7925eV and r0 = 33.875A˚/. The inset table compares Eb obtained respectively from massive Dirac model (MD) and lattice model (LM) for = 2.5.

instead of Φ~k (Eq. (B1)) so that the valley dependence of Kτ is explicit. The interaction potential has the following where σx,y,z is the Pauli matrices for the basis function form at K or K0 point, and ∆ is the energy gap. The angle 2 Φ~k is deﬁned as cot Φ~k = τkx/ky, where τ = ±1 labels 2πe 0 V˜ = F (q), (B4) valley K and K . The conduction and valence band are q q described by spinors:

θ θ where the form factor F (q) = 1/(1 + r0q) modifies the k k −iΦ~k ~ cos( 2 ) ~ sin( 2 )e |c, kiτ = θ iΦ , |v, kiτ = θ , Coulomb interaction in consistency with the real-space sin( k )e ~k − cos( k ) 2 2 interaction potential (Eq. (2)). The BS equation reads: (B2) where angle θ is defined as cos θ = ∆/ε with ε = k k k k X ~ ~ 0 ~ 0 ~ p 2 2 Kτ (k, k )ψ(k ) = Eψ(k). (B5) ∆ + (~vF k) . As discussed in the main text, inter-valley coupling is ~k0 nearly absent for Q~ = 0 excitons. Within each valley, the ∗ kernel of the BS equation in Eq. (5) can be expressed in We can define effective Bohr radius aB, Rydberg en- ∗ terms of band spinors: ergy Ry and fine structure constant α:

~ ~ 0 ~ ~ 0 2 2 2 Kτ (k, k ) =δ~k~k0 Tk − Dτ (k, k ), ∗ 2(~vF ) ∗ 1 e e aB = 2 , Ry = ∗ , α = . (B6) Tk =2εk, e ∆ 2 aB ~vF 1 D (~k,~k0) = V˜ hc,~k|c,~k0i hv,~k0|v,~ki τ A ~k−~k0 τ τ τ τ For notation convenience, we also define parameter β = (α/2)2. 1 ˜ = V~ ~ 0 (1 + cos θk)(1 + cos θk0 ) ∗ ∗ 4A k−k After taking aB as unit of length and Ry as unit of en- ~ ilφ~ iτ(φ~0 −φ~ ) ergy, and using an ansatz ψ(k) = ψ(k)e k , BS equation + 2 sin θk sin θk0 e k k is reduced to the following 1D eigenvalue problem: i2τ(φ~0 −φ~ ) + (1 − cos θk)(1 − cos θk0 )e k k , (B3) Eψ(k) = Tkψ(k) where Tk can be understood as the kinetic energy. Z ∞ ~ ~ 0 0 Dτ (k, k ) is the electron-hole direct interaction, while the − dk (1 + cos θk)(1 + cos θk0 )Ikk0 (l) exchange interaction X (Eq. (5)) is neglected. A is the 0 (B7) +2 sin θ sin θ 0 I 0 (l + τ) area of the 2D system, and φ~k is the orientation angle k k kk ~ 0 0 of k with cot φ~k = kx/ky. In Eq. (B3), angle φ~k is used +(1 − cos θk)(1 − cos θk0 )Ikk0 (l + 2τ) k ψ(k ), 8 where Tk, cos θk and Ikk0 (l) in effective atomic units are: (1)binding energies deviate from the pattern of the 2D hydrogen model; (2)2p states have a larger binding en- 2 p 2 p 2 ergy than 2s; and (3)there is an energy splitting between Tk = 1 + βk , cos θk = 1/ 1 + βk , β l = ±1 2p states within the same valley. Moreover, the 1 Z 2π F (pk2 + k02 − 2kk0 cos φ) cos lφ binding energies calculated by using the Massive Dirac Ikk0 (l) = dφ . model and the lattice model of the main text are close to 4π p 2 02 0 0 k + k − 2kk cos φ each other as shown in the inset table of Fig. 5. (B8) Eq. (B7) makes it clear that within the same valley excitons with quantum number l and −l are not degenerate Finally, we study the massive Dirac model with stan- in energy, because of the wave-vector dependence of the dard Coulomb interaction by taking form factor F (q) to band spinors. However, there is a degeneracy between be 1. In this case, the fine structure constant α controls (τ, l) and (-τ,-l) excitons, which originates from time the deviation of the massive Dirac model from the 2D reversal symmetry. hydrogen model. Fig. 6 presents the binding energy as We apply the massive Dirac model to excitons in A a function of α. In the limit of weak electron-hole in- series of monolayer MoS2. The appropriate parameter teraction (α → 0), the massive Dirac model reduces to 43 values are ~vF = 1.105eV × 3.193A,˚ ∆ = 0.7925eV and the 2D hydrogen model as implied by Eq. (B7) and (B8). ˚ 10 ∗ 4 ∗ r0 = 33.875A/ . The effective atomic units then take Therefore, Eb(1s) = 4Ry and Eb(2s) = Eb(2p) = Ry ∗ ˚ ∗ 2 9 the following value, aB = × 2.18A and Ry = 3.3eV/ , as α goes to 0. 2s and 2p states remain nearly degener- and the fine structure constant α = 4.075/. Eq. (B7) is ate for α < 0.4, and develop prominent energy splitting solved numerically by discretizing the 1D k-space. The at large α. The energy ordering is Eb(2p, l = −1) > result is presented in Fig. 5, which depicts the binding Eb(2s) > Eb(2p, l = +1) in valley K. Note in the case energy of 1s, 2s, and 2p excitons in valley K as a func- of monolayer MoS2 where interaction potential is modi- tion of dielectric constant . It reproduces all essential fied by F (q) (Eq. (B4)), all 2p states have bigger binding features of Q~ = 0 excitons discussed in the main text: energies than 2s states as shown in Fig. 5.

∗ Electronic address: [email protected] Lett. 111, 216805 (2013). † Electronic address: [email protected] 16 K. F. Mak, K. L. McGill, J. Park, and P. L. McEuen, 1 Q. H. Wang, K. Kalantar-Zadeh, A. Kis, J. N. Coleman, Science 344, 1489-1492 (2014). and M. S. Strano, Nature Nanotech. 7, 699-712 (2012). 17 S. Larentis, J. R. Tolsma, B. Fallahazad, D. C. Dillen, K. 2 X. Xu, W. Yao, D. Xiao, and T. F. Heinz, Nature Phys. Kim, A. H. MacDonald, and E. Tutuc, Nano Lett. 14, 10, 343-350 (2014). 2039-2045 (2014). 3 D. Xiao, G.-B. Liu, W. Feng, X. Xu, and W. Yao, Phys. 18 Q. Wang, S. Ge, X. Li , J. Qiu, Y. Ji, J. Feng, and D. Sun, Rev. Lett. 108, 196802 (2012). ACS Nano 7, 11087-11093 (2013). 4 T. Cao, G. Wang, W. Han, H. Ye, C. Zhu, J. Shi, Q. Niu, 19 C. Mai, A. Barrette, Y. Yu, Y. G. Semenov, K. W. Kim, P. Tan, E. Wang, B. Liu, and J. Feng, Nat. Commun. 3, L. Cao, and K. Gundogdu, Nano Lett. 14, 202-206 (2014). 887 (2012). 20 C. Mai, Y. G. Semenov, A. Barrette, Y. Yu, Z. Jin, L. 5 H. Zeng, J. Dai, W. Yao, D. Xiao, and X. Cui, Nature Cao, K. W. Kim, and K. Gundogdu, Phys. Rev. B 90, Nanotech. 7, 490-493 (2012). 041414(R) (2014). 6 K. F. Mak, K.-L. He, J. Shan, and T. F. Heinz, Nature 21 M. M. Glazov, T. Amand, X. Marie, D. Lagarde, L. Bouet, Nanotech. 7, 494-498 (2012). and B. Urbaszek, Phys. Rev. B 89, 201302(R) (2014). 7 A. Splendiani, L. Sun, Y. Zhang, T. Li, J. Kim, C. Chim, 22 T. Yu, and M. W. Wu, Phys. Rev. B 89, 205303 (2014). G. Galli, and F. Wang, Nano Lett. 10, 1271 (2010). 23 C. R. Zhu, K. Zhang, M. Glazov, B. Urbaszek, T. Amand, 8 K. F. Mak, C. Lee, J. Hone, J. Shan, and T. F. Heinz, Z. W. Ji, B. L. Liu, and X. Marie, Phys. Rev. B 90, Phys. Rev. Lett. 105, 136805 (2010). 161302(R) (2014). 9 C.-D. Zhang, A. Johnson, C.-L. Hsu, L.-J. Li, and C.-K. 24 T. C. Berkelbach, M. S. Hybertsen, and D. R. Reichman, Shih, Nano Lett. 14, 2443 (2014). Phys. Rev. B 88, 045318 (2013). 10 C.-J. Zhang, H.-N. Wang, W.-M. Chan, C. Manolatou, and 25 A. Chernikov, T. C. Berkelbach, H. M. Hill, A. Rigosi, Y. F. Rana, Phys. Rev. B 89, 205436 (2014). Li, O. B. Aslan, D. R. Reichman, M. S. Hybertsen, and T. 11 K. F. Mak, K. He, C. Lee, G. H. Lee, J. Hone, T. F. Heinz, F. Heinz, Phys. Rev. Lett. 113, 076802 (2014). and J. Shan, Nature Mater. 12, 207 (2013). 26 L. V. Keldysh, JETP Lett. 29, 658 (1979). 12 J. S. Ross, S. Wu, H. Yu, N. J. Ghimire, A. M. Jones, G. 27 P. Cudazzo, I. V. Tokatly, and A. Rubio, Phys. Rev. B 84, Aivazian, J. Yan, D. G. Mandrus, D. Xiao, W. Yao, and 085406 (2011). X. Xu, Nat. Commun. 4, 1474 (2013). 28 P. Giannozzi, et al., J. Phys.: Condens. Matt. 21, 395502 13 A. Singh, G. Moody, S. Wu, Y. Wu, N. J. Ghimire, J. Yan, (2009). D. G. Mandrus, X. Xu, and X. Li, Phys. Rev. Lett. 112, 29 Under mirror operation M, a vector is transformed as 216804 (2014). M(x, y, z) = (x, y, −z); for a spinor in spin space M|si = 14 A. Ramasubramaniam, Phys. Rev. B 86, 115409 (2012). exp[−iπsz/2]|si with sz being z component Pauli matrix 15 D. Y. Qiu, F. H. da Jornada, and S. G. Louie, Phys. Rev. in spin space. 9

30 H.-Y. Yu, G.-B. Liu, P. Gong, X.-D. Xu, and W. Yao, Nat. 37 Z. Ye, T. Cao, K. OBrien, H. Zhu, X. Yin, Y. Wang, S. G. Commun. 5, 3876 (2014). Louie, and X. Zhang, Nature 513, 214-218 (2014). 31 D. MacNeill, C. Heikes, K. F. Mak, Z. Anderson, A. 38 K. He, N. Kumar, L. Zhao, Z. Wang, K. F. Mak, H. Zhao, Kormányos, V. Zólyomi, J. Park, and D. C. Ralph, Phys. and J. Shan, Phys. Rev. Lett. 113, 026803 (2014). Rev. Lett. 114, 037401 (2015). 39 A. M. Jones, H. Yu, N. J. Ghimire, S. Wu, G. Aivazian, 32 E. J. Sie, J. W. McIver, Y.-H. Lee, L. Fu, J. Kong, and N. J. S. Ross, B. Zhao, J. Yan, D. G. Mandrus, D. Xiao, W. Gedik, arXiv:1407.1825. Yao, and X. Xu, Nature Nanotech. 8, 634-638 (2013). 33 J. Kim, X. Hong, C. Jin, S. Shi, C. S. Chang, M.-H. Chiu, 40 Y. N. Gartstein, X. Li, and C. Zhang, arXiv:1502.00905. L.-J. Li, and F. Wang, Science 346, 1205-1208 (2014). 41 See Ref. 43 for a detailed discussion of conduction bands 34 A. Srivastava, M. Sidler, A. V. Allain, D. S. Lembke, A. energy splitting. Kis, and A. Imamoglu, arXiv:1407.2624. 42 A. Kormányos, V. Zólyomi, N. D. Drummond, P. Rakyta, 35 G. Aivazian, Z. Gong, A. M. Jones, R.-L. Chu, J. Yan, D. G. Burkard, and V. I. Fal’ko, Phys. Rev. B 88, 045416 G Mandrus, C. Zhang, D. Cobden, W. Yao, and X. Xu, (2013). Nature Phys. 11, 148-152 (2015). 43 G.-B. Liu, W.-Y. Shan, Y.-G. Yao, W. Yao, and D. Xiao, 36 X. Li, T. Cao, Q. Niu, J. Shi, and J Feng, PNAS 110, 3738 Phys. Rev. B 88, 085433 (2013). (2013).