Arxiv:2006.00399V2 [Cond-Mat.Mes-Hall] 8 Oct 2020

Twisted Trilayer Graphene: a Precisely Tunable Platform for Correlated Electrons

Ziyan Zhu,1 Stephen Carr,1 Daniel Massatt,2 Mitchell Luskin,3 and Efthimios Kaxiras1, 4 1Department of Physics, Harvard University, Cambridge, Massachusetts 02138, USA 2Department of Statistics, The University of Chicago, Chicago, Illinois 60637, USA 3School of Mathematics, University of Minnesota - Twin Cities, Minneapolis, Minnesota 55455, USA 4John A. Paulson School of Engineering and Applied Sciences, Harvard University, Cambridge, Massachusetts 02138, USA We introduce twisted trilayer graphene (tTLG) with two independent twist angles as an ideal system for the precise tuning of the electronic interlayer coupling to maximize the effect of correlated behaviors. As established by experiment and theory in the related twisted bilayer graphene system, van Hove singularities (VHS) in the density of states can be used as a proxy of the tendency for correlated behaviors. To explore the evolution of VHS in the twist-angle phase space of tTLG, we present a general low-energy electronic structure model for any pair of twist angles. We show that the basis of the model has infinite dimensions even at a finite energy cutoff and that no Brillouin zone exists even in the continuum limit. Using this model, we demonstrate that the tTLG system exhibits a wide range of magic angles at which VHS merge and the density of states has a sharp peak at the charge-neutrality point through two distinct mechanisms: the incommensurate perturbation of twisted bilayer graphene’s flatbands or the equal hybridization between two bilayer moiré superlattices.

Introduction. — Electronic properties in stacked of moiré length as a function of twist angles, in which graphene layers can be tuned by a small twist angle each lobe corresponds to the region where a different that modifies the interlayer interaction strength, an ef- harmonic (m, n) dominates. The moiré of moiré pat- fect referred to as “twistronics” [1]. As the twist angle terns can be discerned visually only near the (N, 1) or approaches a critical “magic angle” ( 1.05◦ in twisted (1,N) lobes for N Z. Generally, multiple harmonics ∼ ∈ bilayer graphene), the two van Hove singularities (VHS) have competing length scales (see Supplemental Material in the density of states (DOS) of each monolayer merge, Sec. I [22]). Therefore, tTLG cannot be approximated resulting in a sharp peak associated with flatbands, lead- by two aligned tBLG and a general expression for the ing to the emergence of strongly correlated electronic trilayer supercell does not exist, making it fundamen- phases [2]. The small twist angle gives rise to large-scale tally different than multilayered vdW heterostructures repeating patterns, known as moiré patterns. Unconven- with a single twist angle [23–26]. The lack of a super- tional correlated states have now been observed in many cell approximation and the large length scale pose many van der Waals (vdW) heterostructures with one twist an- computational challenges to the theoretical modeling of gle, e.g., twisted bilayer graphene (tBLG) and twisted tTLG. While there have been some theoretical studies double bilayer graphene [3–15]. In these systems, elec- of tTLG [19, 27, 28], including an accurate treatment of trons responsible for the correlation effects localize at the any twist angles by Amorim and Castro [27], an elec- moiré scale [16–18]. tronic structure model incorporating both accuracy and efficiency is lacking; this severely restricts our ability to The addition of a third layer introduces a new degree investigate its electronic properties and the potential for of freedom, the second twist angle, allowing for the fur- correlated phases, which have been observed recently in ther tuning of electron correlations. In twisted trilayer tTLG at the moiré of moiré scale [29]. graphene (tTLG) with two consecutive twist angles, θ12 and θ23, the beating of two bilayer moiré patterns leads to Here, we present tTLG as a platform to precisely higher-order patterns (moiré of moiré). The length scale tune twistronic correlations, using the VHS intensity as of these is orders of magnitude larger than the bilayer a proxy for strong correlations. We derive a general moiré [Fig.1(a)] [19–21]. Unlike in tBLG where only the momentum-space model to study the electronic proper- arXiv:2006.00399v2 [cond-mat.mes-hall] 8 Oct 2020 lowest-order moiré pattern dominates in the continuum ties of the two-independent-twist-angle tTLG system us- limit, the dominant harmonic is twist-angle dependent in ing a low-energy k p model that provides computational · tTLG. For a given moiré of moiré harmonic labeled by efficiency and removes the constraint on the twist an- (m, n), the primitive reciprocal lattice vectors are given gle in atomistic calculations with supercells. Using this as the column vectors of GH = mG nG , where model, we explore the tTLG phase space. We find that mn 12 − 23 the matrix Gij spans the bilayer moiré reciprocal space the two bilayer moiré superlattices hybridize when the between layers i (Li) and j (Lj). The real space moiré of two twist angles are equal, minimizing the separation H H 1 H T moiré supercell Amn is obtained by Amn = 2π (Gmn)− , between the two lowest VHS at a critical angle. At gen- with the norm of its column vectors being the moiré eral twist angles, there exists a wide range of values at of moiré length. Figure1b shows the dominant moiré which the VHS merge and the DOS is sharply peaked at 2 the charge-neutrality point (CNP). These magic angles where Γ is the monolayer unit cell area, τ (τ ) is the | | α β can be understood as a tBLG magic angle modified by position of the sublattice α (β), G(`) is a reciprocal space ˜ij an incommensurate perturbative potential from the third lattice vector in L`, and tαβ(p) is the momentum-space layer. Our analysis is well suited to guide experimental hopping parameter between sublattice α in Li and sub- searches for correlation effects and enables the interpre- lattice β in Lj. The δ function imposes the constraint tation of otherwise unclear experimental findings [29]. on the values of k(`), dictating the interlayer scattering selection rule. The above expressions are equivalent to a real-space tight-binding model in the Bloch basis (See Supplemental Material, Sec. IIA, for derivation [22]). Unlike tBLG [31–35], the momentum-space basis in tTLG is infinitely dimensional and lacks a Brillouin zone even in the continuum limit. In bilayers, coupled momentum states satisfy the selection rule k(1) k(2) = G(1) (2) (1) − (1) (1)− G [33]. Note that for a given G = mb1 + nb2 (2) (2) (2) for m, n Z, we also have G = mb1 + nb2 for the ∈ (`) same m, n, where bi is the i-th component of the primitive reciprocal lattice vector of L`, since other hopping processes are much higher in energy. As G(`) increases, | | the scattered momentum, k0, becomes farther away from FIG. 1: Illustration of moiré of moiré pattern in tTLG for the Dirac point. Therefore, to implement a finite cutoff, θ12 = 2.6◦, θ23 = 2.8◦. Red and blue points represent the lattice points of the bilayer moiré supercells between L1-L2 we can simply constrain the magnitude of the scattered (`) and L2-L3 respectively. Black arrows represent dominant momentum k0 = G for ` = 1, 2. Physically, k0 is a moiré of moiré supercell lattice vectors. A blowup of the monolayer reciprocal lattice vector that can scatter to a small boxed area is shown below, with points representing nearby momentum in the other layer. In contrast, in tri- monolayer lattice points, for L1 and L2 on the left half and layers, the momentum states that form the basis of the for L2 and L3 on the right half. The moiré lattice vectors Hamiltonian are connected in a more complicated way. (red and blue arrows) are slightly rotated and have different A given k(1) can couple to a momentum state k(2) that lattice constants. (b) The dominant moiré of moiré length (2) (1) (2) (1) on a logarithmic color scale. The black star corresponds to satisfies k = k + G G , same as in bilayers. (2) − (3) the twist angle in (a), and (m, n) labels the dominant moiré Each k can then couple to a momentum state k of moiré harmonic in the nearby lobe. Black dashed line through the second selection rule [Eq. (2)], resulting in represents θ12 = θ23. the following final momentum:

(3) (1) (2) (1) (3) (2) k = k + [G G ] + [G G0 ], (3) Momentum-space Hamiltonian. — To obtain the − − electronic structure model for tTLG, we employ a where the reciprocal lattice vectors satisfy G(2) G(1) = momentum-space method by taking the Fourier trans- (12) (12) (3) (2) (23) − (23) mb1 + nb2 and G G0 = m0b1 + n0b2 for form of the real-space tight-binding model. At a momen- (ij−) (j) (i) m, n, n0, m0 Z, with b = b b being the bi- tum k (referred to as the center site), the model can be ∈ k k − k layer moiré reciprocal space lattice vectors. Equation (3) formally represented by a 3 3 block matrix: × suggests that L1 and L3 are coupled through L2, even H1(k) T 12 0 though a direct interlayer hopping is not allowed. Unlike 12 2 23 the simple 2D momentum crystal in bilayers, here the in- (k) = T † H (k) T . (1) (12) (23) H  23 3  (3) 0 T † H (k) commensuration between bk and bk creates for k   a 4D structure that is projected onto 2D. The diagonal blocks are the monolayer graphene tight- Equation (3) suggests that in L` of tTLG, k0 is given (i) (j) binding Hamiltonians in the rotated basis [30], represent- by k0 = G + G for ` = i, j. To implement a cut- 6 ing the intralayer hopping. The off-diagonal blocks rep- off, we should impose k0 k for some cutoff value | | ≤ c resent the interlayer hopping. The interlayer terms that kc. However, the incommensurability of twisted trilayers (i) (j) connect two momentum degrees of freedom k and k suggests that k0 can be arbitrarily small and imposing | | in Li and Lj are given as, k0 k still leads to an infinite basis. For example, in | | ≤ c Fig.2(a), even though G(2) lies outside of the cutoff, the ij (i) (j) 1 iG(i) τ (i) T [k , k ] = e · α resulting k0 is still a relevant low-energy degree of free- αβ Γ | | G(i),G(j) dom, due to the two-step scattering process. A similar X (2) ij iG(j) τ (i) construction can be made for all other G outside of the ˜ (i) (i) β tαβ[k + k + G ]e− · δk(i) G(i),k(j) G(j) , × − − cutoff radius, which means within a finite cutoff, there are (2) infinitely many coupled momentum states. In practice, 3 another set of cutoff conditions needs to be implemented, (`) (`) namely G k . With the constraint on G , the k0 | | ≤ c | | in Fig.2(a) is no longer allowed. In this way, we ignore (`) the cases where G is large but k0 is small, leading | | | | to the neglect of some low-energy degrees of freedom and hence convergence is not guaranteed, which merits future work (see Supplemental Material, Sec. IID, for convergence study [22]). In this work, we choose k = 4 b(`) , c | | with 5600 momenta, such that the properties of inter- ∼ est (e.g., DOS maximum and the VHS location) do not change significantly as kc increases. FIG. 2: Momentum degrees of freedom for tTLG at We take the low-energy limit by expanding around the θ12 = 2.2◦, θ23 = 2◦. Red, blue, and green are the reciprocal (`) (`) Dirac point of each layer, KL`, letting k = q + KL`, lattice vectors of L1, L2, and L3 respectively. The origin is which simplifies the model proposed by Amorim and Cas- the Dirac point of L2. (a) Extended zone scheme, with the (2) (3) tro [27]. The intralayer Hamiltonian becomes the rotated orange circle indicating the cutoff in k0 . k0 = G + G ` θ` θ` θ` 1 | | (2) Dirac equation, H = vF q (σx , σy ), where σx = falls within the cutoff radius 10Å− despite both G and · − (3) | | (2) θ` G being large. The momenta of L3 are centered at G . σx cos θ` σy sin θ` and σy = σx sin θ` + σy cos θ` are ro- | | − θ = θ , θ = 0, θ = θ (b) Reduced zone scheme, folded back to the monolayer tated Pauli matrices with 1 12 2 3 23, (`) (`) 6 − Dirac points, q = k KL`. This basis corresponds to vF = 0.8 10 cm/s is the Fermi velocity [36], and q = − (`) × the same twist angle as (a) but with an additional constraint k + k KL`. For the interlayer hopping, we make the (`) (`) − ˜ij (i) (i) ˜ij (i) G kc = 6 b , leading to 26 921 momenta. approximation that tαβ[k +k +G ] tαβ[G +KLi] | | ≤ | | (i) (i) ≈ since k , q KLi , G , for k near the Dirac point. | | | | | | | t˜ij |(p) p Due to the rapid decay of αβ as increases [2, 36, 37], using the same constant for the DOS of the full Hamil- we keep only the first shell in the summation in Eq. (2): tonian. We adapt the Gaussian FWHM, κ, based on the twist angle θ : for θ 2◦, κ = 0.35 meV; 3 `,`+1 `,`+1 ≤ ij (i) (j) ij θ (2◦, 3.9◦], κ = 1.2 meV θ > 3.9◦ T [q , q ] = T δ (i) (j) ij , (4) for `,`+1 ; for `,`+1 , αβ n,αβ q q , qn ∈ n=1 − − κ = 2.4 meV. X Evolution of VHS. — We explore next the behavior of ij ij 1 ij where q = KLi KLj, q = − (2π/3)q , and VHS as a function of twist angles in tTLG, by investi- 1 − 2 R 1 qij = (2π/3)qij using a counterclockwise rotation ma- gating the DOS enhancement and the narrowing of the 3 R 1 trix (θ). We include out-of-plane relaxation by letting separation between VHS (referred to as the VHS gap). ij R ij ij ij tAA = tBB = ω0 = 0.07 eV and tAB = tBA = ω1 = We define a magic angle approximately as a geometry 0.11 eV [32, 38]. In the matrix form, where both features are achieved. Figure3(a) shows the DOS of tTLG at θ12 = θ23. The bright regions repre- ω ω ω ω φ¯ T ij = 0 1 ,T ij = 0 1 ,T ij = T¯ij, (5) sent VHS. As the twist angle decreases, the VHS gap 1 ω ω 2 ω φ ω 3 2 1 0 1 0 first decreases and then increases after reaching a minimum at 2.1◦. This behavior is similar to the evolution ∼ where φ = exp(i2π/3) and z¯ indicates the complex con- of VHS in tBLG in which changing the twist angle tunes jugate of z. In tBLG, with the low-energy expansion, the hybridization between two monolayer Dirac cones. In momenta q(1) and q(2) form a hexagonal lattice with the tTLG with θ12 = θ23, varying the twist angle changes the neighboring hexagon corners representing states from al- hybridization strength between the two identical bilayer ternating layers (a moiré momentum lattice) [2, 33]. In moiré superlattices. However, the two VHS can never tTLG, on top of each lattice point of the L1-L2 moiré mo- merge at the CNP, with the minimum VHS gap being mentum lattice, the additional scattering process creates 20 meV at 2.1◦. The DOS is also orders of magnitude ∼ a copy of the L2-L3 moiré momentum lattice (Fig.2b), lower than at the tBLG magic angle. For general twist suggesting the absence of a Brillouin zone. angles, Fig.3(b) shows the DOS as a function of θ12 with Density of states — We use Gaussian smearing to θ23 = 3◦. Unlike when θ12 = θ23, the two VHS approach obtain the total DOS, summing over the two bilayer each other as the twist angle decreases and merge when moiré Brillouin zones, each discretized using a 22 22 1.3◦ θ12 1.6◦, resulting in a sharp DOS peak. × ≤ ≤ grid [39] (see Supplemental Material Sec. IIC for the ex- To investigate the nature of DOS enhancements in pression [22]). For normalization, we first calculate the tTLG, we performed calculations over an entire region DOS of only the intralayer Hamiltonian, which reduces of the θ12, θ23 parameter space. Figure4 shows the DOS to three independent copies of monolayer graphene [30]. maximum and the VHS gap, ∆E, as a function of both We then obtain the normalization constant by fixing the twist angles [40]. The magic-angle condition is met at a prefactor to the expected low-energy monolayer DOS and wide range of twist angles that follows a smooth curve but 4

for example, when θ23 , α23 0, Eq. (7) becomes 2 1 → ∞ → α12 = 3 , which is the tBLG magic-angle condition [2].

FIG. 3: (a) DOS as a function of twist angle for θ12 = θ23. (b) DOS as a function θ12 at θ23 = 3◦, both on a logarithmic color scale. disappears near the diagonal. Although there is no signif- FIG. 4: (a) DOS maximum and (b) VHS gap, ∆E, as a icant DOS enhancement at θ12 = θ23, the DOS maximum function of twist angles on a logarithmic color scale. The is higher compared to the nearby regions where θ and 12 black solid line follows the tTLG magic angles predicted by θ23 differ slightly [light yellow region within the dotted Eq. (7). Vertical and horizontal black dashed lines lines in Fig.4(a)]. correspond to θ12 and θ23 at the tBLG magic angle We now examine the magic angles away from the di- respectively. Within the dotted lines is roughly the region agonal. In the limit where θ θ or θ θ , that can be understood as the hybridization between two 12 23 12 23 tTLG decomposes into a decoupled tBLG moiré supercell bilayer moiré superlattices. and a graphene monolayer; the monolayer does not contribute significantly to the low-energy features. There- The evolution of VHS along the diagonal likely has fore, we observe that the tTLG magic-angle curve asymp- a different origin than the magic angles for θ12 = θ23. 6 totically approaches the tBLG magic angle [dashed lines Perturbation theory predicts that vF∗ can reach 0 at in Fig.4(a)] for large θ12 or θ23. We verified numerically θ12 = θ23 = 1.72◦. In the numerical calculations, how- that when one twist angle is very large, θ12 = 40◦ for ever, we do not observe vF∗ = 0 at equal twist angles, and instance, the DOS maximum occurs exactly when θ23 is the twist angle with the minimal VHS gap (2.1◦) devi- at the tBLG magic angle. The continuous curve and its ates from the perturbation theory prediction. The dis- asymptotic behavior suggest that these magic angles can crepancy suggests that the perturbation argument does be understood as the magic-angle tBLG modified by an not apply to equal twist angles, since features near the effective potential, V , from the third layer. We can quali- diagonal are more aptly described by the hybridization tatively analyze this argument using perturbation theory, between the two bilayer moiré superlattices with a shared by truncating the momentum space to the first shell, in- middle layer, rather than between two independent unit cluding one state from L2 and three states each from L1 cells as in tBLG. and L3. We obtain the renormalized Fermi velocity vF∗ Moiré of moiré. — In magic-angle tBLG, correlated by extracting the coefficient of the first-order effective states occur at the half-filling of the moiré supercell by Hamiltonian in q in the form of a Dirac Hamiltonian, filling two isolated flatbands [3,4, 16–18]. In tTLG, given by even though the origin of some magic angles is per- turbed tBLG, filling each flatband corresponds to filling 2 2 1 3(α12 + α23) the moiré of moiré supercell rather than the bilayer moiré vF∗ = − 2 2 vF , (6) 1 + 6(α12 + α23) cell because the incommensurate effective potential modifies the relevant supercell area. α = ω/(v k ) k = 8π sin(θ /2)/(3a ) where ij F θij , θij ij G , as- We compare our results to a simplified model that ap- suming that ω0 = ω1 = ω. The Hamiltonian and its proximates tTLG as two aligned moiré cells [28]. While derivation are provided in Section III of the Supplemen- we observe similar qualitative behaviors, the simplified tal Material [22]. Magic angles occur when vF∗ vanishes, model fails to capture physics at the moiré of moiré scale leading to the following condition: and does not predict as drastic a DOS enhancement as 1 our work. Moreover, the simplified model requires a new α2 + α2 = . (7) 12 23 3 basis for different sets of twist angles, making it difficult to generalize – limitations that our model overcomes. We The solid line in Fig.4(a) corresponds to θ12 and θ23 that include a comparison between the two models in Sec- satisfy Eq. (7), which matches the DOS peaks and ∆E tion IV of the Supplemental Material IV [22]. minima in Fig.4(a), (b). Taking the large angle limit, In summary, we explore the rich electronic behavior 5 of tTLG in its twist-angle phase space. We offer a gen- moiré superlattices with pressure,” Nature (London) 557, eral low-energy momentum-space model to obtain elec- 404–408 (2018). tronic structure in tTLG. We show that the twisted tri- [7] Matthew Yankowitz, Shaowen Chen, Hryhoriy Polshyn, layer momentum-space model does not have a Brillouin Yuxuan Zhang, K. Watanabe, T. Taniguchi, David Graf, Andrea F. Young, and Cory R. Dean, “Tuning super- zone and has an infinitely sized basis. Although we do conductivity in twisted bilayer graphene,” Science 363, not predict correlation strengths directly, we can use the 1059–1064 (2019). presence of VHS as a proxy for electronic correlation. [8] Cheng Shen, Na Li, Shuopei Wang, Yanchong Zhao, We show that the tTLG system exhibits a wide range of Jian Tang, Jieying Liu, Jinpeng Tian, Yanbang Chu, magic angles with merging VHS at the CNP. Away from Kenji Watanabe, Takashi Taniguchi, Rong Yang, Zi Yang equal twist angles, the origin of the magic angles can be Meng, Dongxia Shi, and Guangyu Zhang, “Corre- understood as tBLG in an incommensurate perturbative lated states in twisted double bilayer graphene,” Nature Physics 16, 520–525 (2020). potential. At equal twist angles, the electronic proper- [9] Xiaomeng Liu, Zeyu Hao, Eslam Khalaf, Jong Yeon ties are a result of the hybridization between two bilayer Lee, Yuval Ronen, Hyobin Yoo, Danial Haei Najafabadi, moiré superlattices that share the middle layer. Tun- Kenji Watanabe, Takashi Taniguchi, Ashvin Vishwanath, ing the twist angle makes it possible to traverse between et al., Nature 583, 221–225 (2020). these two regimes. Our MATLAB code for the model is [10] Yuan Cao, Daniel Rodan-Legrain, Oriol Rubies-Bigorda, openly available [41]. Jeong Min Park, Kenji Watanabe, Takashi Taniguchi, and Pablo Jarillo-Herrero, “Electric Field Tunable Cor- We thank Ke Wang, Xi Zhang, Kan-Ting Tsai, Fran- related States and Magnetic Phase Transitions in Twisted cisco Guinea, Philip Kim, and Paul Cazeaux for help- Bilayer-Bilayer Graphene,” Nature , 1–6 (2020). ful discussions. This work was supported by the STC [11] G. William Burg, Jihang Zhu, Takashi Taniguchi, Kenji Center for Integrated Quantum Materials, NSF Grant Watanabe, Allan H. MacDonald, and Emanuel Tutuc, No. DMR-1231319, ARO MURI Grant No. W911NF- “Correlated insulating states in twisted double bilayer 14-0247, and NSF DMREF Grant No. 1922165. Cal- graphene,” Phys. Rev. Lett. 123, 197702 (2019). culations were performed on the Odyssey cluster sup- [12] Lei Wang, En-Min Shih, Augusto Ghiotto, Lede Xian, Daniel A. Rhodes, Cheng Tan, Martin Claassen, ported by the FAS Division of Science, Research Com- Dante M. Kennes, Yusong Bai, Bumho Kim, Kenji puting Group at Harvard University. Watanabe, Takashi Taniguchi, Xiaoyang Zhu, James Hone, Angel Rubio, Abhay Pasupathy, and Cory R. Dean, “Magic continuum in twisted bilayer WSe2,” arXiv e-prints , arXiv:1910.12147 (2019), arXiv:1910.12147 [cond-mat.mes-hall]. [1] Stephen Carr, Daniel Massatt, Shiang Fang, Paul [13] Yu Saito, Jingyuan Ge, Kenji Watanabe, Takashi Cazeaux, Mitchell Luskin, and Efthimios Kaxiras, Taniguchi, and Andrea F. Young, “Decoupling super- “Twistronics: Manipulating the electronic properties of conductivity and correlated insulators in twisted bilayer two-dimensional layered structures through their twist graphene,” arXiv e-prints , arXiv:1911.13302 (2019), angle,” Phys. Rev. B 95, 075420 (2017). arXiv:1911.13302 [cond-mat.mes-hall]. [2] Rafi Bistritzer and Allan H. MacDonald, “Moiré bands [14] Zhiming Zhang, Rachel Myers, Kenji Watanabe, Takashi in twisted double-layer graphene,” Proceedings of the Na- Taniguchi, and Brian J. LeRoy, “Probing the wavefunc- tional Academy of Science 108, 12233–12237 (2011). tions of correlated states in magic angle graphene,” arXiv [3] Yuan Cao, Valla Fatemi, Ahmet Demir, Shiang Fang, e-prints , arXiv:2003.09482 (2020), arXiv:2003.09482 Spencer L. Tomarken, Jason Y. Luo, Javier D. [cond-mat.mes-hall]. Sanchez-Yamagishi, Kenji Watanabe, Takashi Taniguchi, [15] Xiaoxue Liu, Zhi Wang, K. Watanabe, T. Taniguchi, Efthimios Kaxiras, Ray C. Ashoori, and Pablo Jarillo- Oskar Vafek, and J. I. A. Li, “Tuning electron cor- Herrero, “Correlated insulator behaviour at half-filling relation in magic-angle twisted bilayer graphene using in magic-angle graphene superlattices,” Nature (London) Coulomb screening,” arXiv e-prints , arXiv:2003.11072 556, 80–84 (2018). (2020), arXiv:2003.11072 [cond-mat.mes-hall]. [4] Yuan Cao, Valla Fatemi, Shiang Fang, Kenji Watanabe, [16] Mikito Koshino, Noah F. Q. Yuan, Takashi Koretsune, Takashi Taniguchi, Efthimios Kaxiras, and Pablo Jarillo- Masayuki Ochi, Kazuhiko Kuroki, and Liang Fu, “Max- Herrero, “Unconventional superconductivity in magic- imally Localized Wannier Orbitals and the Extended angle graphene superlattices,” Nature (London) 556, 43– Hubbard Model for Twisted Bilayer Graphene,” Physi- 50 (2018). cal Review X 8, 031087 (2018). [5] Guorui Chen, Lili Jiang, Shuang Wu, Bosai Lyu, [17] Hoi Chun Po, Liujun Zou, T. Senthil, and Ashvin Vish- Hongyuan Li, Bheema Lingam Chittari, Kenji Watan- wanath, “Faithful tight-binding models and fragile topol- abe, Takashi Taniguchi, Zhiwen Shi, Jeil Jung, Yuanbo ogy of magic-angle bilayer graphene,” Phys. Rev. B 99, Zhang, and Feng Wang, “Evidence of a gate-tunable 195455 (2019). Mott insulator in a trilayer graphene moiré superlattice,” [18] Stephen Carr, Shiang Fang, Hoi Chun Po, Ashvin Vish- Nature Physics 15, 237–241 (2019). wanath, and Efthimios Kaxiras, “Derivation of Wannier [6] Matthew Yankowitz, Jeil Jung, Evan Laksono, Nico- orbitals and minimal-basis tight-binding Hamiltonians las Leconte, Bheema L. Chittari, K. Watanabe, for twisted bilayer graphene: First-principles approach,” T. Taniguchi, Shaffique Adam, David Graf, and Cory R. Physical Review Research 1, 033072 (2019). Dean, “Dynamic band-structure tuning of graphene [19] Ziyan Zhu, Paul Cazeaux, Mitchell Luskin, and 6

Efthimios Kaxiras, “Modeling mechanical relaxation in laxation and energy band modulation in twisted bilayer incommensurate trilayer van der waals heterostructures,” graphene,” Phys. Rev. B 96, 075311 (2017). Phys. Rev. B 101, 224107 (2020). [32] Stephen Carr, Shiang Fang, Ziyan Zhu, and Efthimios [20] M. Anđelković, S. P. Milovanović, L. Covaci, and F. M. Kaxiras, “Exact continuum model for low-energy elec- Peeters, “Double moiré with a twist: super-moiré in en- tronic states of twisted bilayer graphene,” Physical Re- capsulated graphene,” Nano Letters 20, 979–988 (2020). view Research 1, 013001 (2019). [21] N Leconte and J Jung, “Commensurate and incommensu- [33] Shiang Fang, Stephen Carr, Ziyan Zhu, Daniel Mas- rate double moire interference in graphene encapsulated satt, and Efthimios Kaxiras, “Angle-Dependent Ab ini- by hexagonal boron nitride,” 2D Materials 7, 031005 tio Low-Energy Hamiltonians for a Relaxed Twisted (2020). Bilayer Graphene Heterostructure,” arXiv e-prints [22] See Supplemental Material for the calculation of moiré , arXiv:1908.00058 (2019), arXiv:1908.00058 [cond- of moiré length, a detailed derivation of the momentum mat.mes-hall]. space model and convergence test, the derivation of the [34] Francisco Guinea and Niels R. Walet, “Continuum mod- analytical expression of the tTLG magic angles, and com- els for twisted bilayer graphene: Effect of lattice deforma- parisons with other models, which includes Refs. [42, 43]. tion and hopping parameters,” Phys. Rev. B 99, 205134 [23] Xiao Li, Fengcheng Wu, and Allan H. MacDon- (2019). ald, “Electronic Structure of Single-Twist Trilayer [35] Nicolas Leconte, Srivani Javvaji, Jiaqi An, and Jeil Graphene,” arXiv e-prints , arXiv:1907.12338 (2019), Jung, “Relaxation Effects in Twisted Bilayer Graphene: a arXiv:1907.12338 [cond-mat.mtrl-sci]. Multi-Scale Approach,” arXiv e-prints , arXiv:1910.12805 [24] Stephen Carr, Chenyuan Li, Ziyan Zhu, Efthimios Kaxi- (2019), arXiv:1910.12805 [cond-mat.mes-hall]. ras, Subir Sachdev, and Alexander Kruchkov, “Ultra- [36] Shiang Fang and Efthimios Kaxiras, “Electronic struc- heavy and ultrarelativistic dirac quasiparticles in sand- ture theory of weakly interacting bilayers,” Phys. Rev. B wiched graphenes,” Nano Letters 20, 3030–3038 (2020). 93, 235153 (2016). [25] Shaowen Chen, Minhao He, Ya-Hui Zhang, Valerie [37] Gonçalo Catarina, Bruno Amorim, Eduardo V Cas- Hsieh, Zaiyao Fei, K. Watanabe, T. Taniguchi, David H. tro, João MVP Lopes, and Nuno Peres, “Twisted bi- Cobden, Xiaodong Xu, Cory R. Dean, and Matthew layer graphene: Low-energy physics, electronic and op- Yankowitz, “Electrically tunable correlated and topolog- tical properties,” Handbook of Graphene, Volume 3: ical states in twisted monolayer-bilayer graphene,” arXiv Graphene-like 2D Materials , 177 (2019). e-prints , arXiv:2004.11340 (2020), arXiv:2004.11340 [38] Nguyen N. T. Nam and Mikito Koshino, “Lattice re- [cond-mat.mes-hall]. laxation and energy band modulation in twisted bilayer [26] Youngju Park, Bheema Lingam Chittari, and Jeil graphene,” Phys. Rev. B 96, 075311 (2017). Jung, “Gate-tunable topological flat bands in twisted [39] Daniel Massatt, Stephen Carr, Mitchell Luskin, and monolayer-bilayer graphene,” Phys. Rev. B 102, 035411 Christoph Ortner, “Incommensurate heterostructures in (2020). momentum space,” Multiscale Modeling & Simulation [27] B. Amorim and Eduardo V. Castro, “Electronic spec- 16, 429–451 (2018). tral properties of incommensurate twisted trilayer [40] The data set consists of a 43 43 sampling. A Gaussian × graphene,” arXiv e-prints , arXiv:1807.11909 (2018), convolution kernel is applied for smoothening. arXiv:1807.11909 [cond-mat.mes-hall]. [41] Ziyan Zhu, Carr Stephen, Massatt Daniel, Luskin [28] Christophe Mora, Nicolas Regnault, and B. Andrei Mitchell, and Kaxiras Efthimios, “Model for Twisted Bernevig, “Flatbands and Perfect Metal in Trilayer Moiré Trilayer Graphene: a precisely tunable platform for cor- Graphene,” Phys. Rev. Lett. 123, 026402 (2019). related electrons: https://github.com/ziyanzzhu/ttlg,” [29] Kan-Ting Tsai, Xi Zhang, Ziyan Zhu, Yujie Luo, (2020). Stephen Carr, Mitchell Luskin, Efthimios Kaxiras, and [42] Matthew Yankowitz, Jiamin Xue, Daniel Cor- Ke Wang, “Correlated Superconducting and Insulating mode, Javier D. Sanchez-Yamagishi, K. Watanabe, States in Twisted Trilayer Graphene Moiré of Moiré Su- T. Taniguchi, Pablo Jarillo-Herrero, Philippe Jacquod, perlattices,” arXiv e-prints , arXiv:1912.03375 (2019), and Brian J. Leroy, “Emergence of superlattice Dirac arXiv:1912.03375 [cond-mat.mes-hall]. points in graphene on hexagonal boron nitride,” Nature [30] A. H. Castro Neto, F. Guinea, N. M. R. Peres, K. S. Physics 8, 382–386 (2012). Novoselov, and A. K. Geim, “The electronic properties [43] Wei-Jie Zuo, Jia-Bin Qiao, Dong-Lin Ma, Long-Jing Yin, of graphene,” Reviews of Modern Physics 81, 109–162 Gan Sun, Jun-Yang Zhang, Li-Yang Guan, and Lin (2009). He, “Scanning tunneling microscopy and spectroscopy of [31] Nguyen N. T. Nam and Mikito Koshino, “Lattice re- twisted trilayer graphene,” Physical Review B 97, 035440 (2018). Supplemental Material for “Twisted Trilayer Graphene: a Precisely Tunable Platform for Correlated Electrons”

The Supplemental Material includes four sections. In SectionI, we discuss the geometry of the twisted trilayer graphene (tTLG) and calculate the higher-order moiré of moiré lengths. In SectionII, we present a detailed derivation of the momentum-space model and test its convergence. In SectionIII, we derive analytically the magic angles in tTLG. Finally, in SectionIV, we compare and contrast our results with a simpliﬁed model proposed by Mora et al. [S1] as well as results obtained with a full model without the low-energy expansion proposed by Amorim and Castro [S2].

I. CALCULATION OF MOIRÉ OF MOIRÉ LENGTHS

The atomic and reciprocal space geometry of tTLG with two independent twist angles are shown in Fig. S1(a). The monolayer lattice vectors are deﬁned as the column vectors of the following matrix:

1 1/2 A = a = a1 a2 , (S1) 0 G 0 √3/2 where aG = 2.4768 Å is the graphene lattice constant (as obtained from DFT). The l-th layer will be referred to as L`. We assume that L2 is unrotated, with L1 rotated clockwise by θ12 and L3 rotated counterclockwise by θ23. Deﬁning the counterclockwise rotation matrix cos θ sin θ (θ) = − , (S2) R sin θ cos θ

the lattice vectors of the three layers can be written as A1 = ( θ12)A0, A2 = A0, and A3 = (θ23)A0 respectively, (`) R − R with the column vectors denoted as ai for ` = 1, 2, 3 and i = 1, 2. The monolayer reciprocal lattice vectors are given T (2) by the columns of G` = 2πA`− . For example, the reciprocal lattice vectors of L2 are b1 = 2π/aG(1, √3/3) and (2) (2) (1) − b2 = 2π/aG(0, 2/√3). The K point of L2 is given as KL2 = (2b + b )/3 = (4π/(3aG), 0). The reciprocal lattice (i) vectors of layers L1 and L3 can be obtained by acting ( θ12) and (θ23) on b for i = 1, 2. We also denote the (`) R − R (`) monolayer unit cell of layer ` to be Γ and the reciprocal space to be Γ ∗. The twisted trilayer system exhibits higher order moiré of moiré patterns due to the interference between the two ij bilayer moiré patterns. To the lowest order, the bilayer moiré length between layers i and j is given by aM = aG/ sin θij. We denote the bilayer moiré superlattice between layers ij to be Γij, spanned by the column vectors of matrix (ij) (ij) Aij = (a1 a2 ). The bilayer moiré Brillouin zone between layers i and j are given by the column vectors of T T T Gij = Gj Gi = 2π(Aj− Ai− ). The lattice vector of the moiré superlattice is the column vectors of Aij = 2πGij− . − − 12 23 After some algebra, we obtain the lattice vectors for the two bilayer supercells, ai , ai for i = 1, 2:

arXiv:2006.00399v2 [cond-mat.mes-hall] 8 Oct 2020 a a12 = G [(1 cos θ )ˆx sin θ yˆ] , 1 2(1 cos θ ) − 12 − 12 − 12 a 1 cos θ √3 1 √3 a12 = G − 12 + sin θ xˆ + sin θ + (1 cos θ ) yˆ , 2 2(1 cos θ ) 2 2 12 −2 12 2 − 12 − 12 " ! ! # a a23 = G [(1 cos θ )ˆx + sin θ yˆ] , 1 2(1 cos θ ) − 23 23 − 23 aG 1 cos θ23 √3 1 √3 a23 = − sin θ xˆ + sin θ + (1 cos θ ) yˆ . (S3) 2 2(1 cos θ ) 2 − 2 23 2 23 2 − 23 − 23 " ! ! # 2

FIG. S1: Lattice geometries of the tTLG system. (a) The twisted trilayer graphene system in real space (left) and momentum space with their original monolayer reciprocal lattice vectors (right). (b) Bilayer moiré Brillouin zone for θ12 = θ23 with high symmetry points

Note that there is a small twist angle between the two bilayer moiré superlattices. Moreover, for general twist angles θ = θ , the two bilayer moiré cells have a lattice mismatch. The twist angle φ and the lattice mismatch δ between 12 6 23 the two bilayer moiré patterns give rise to higher-order moiré of moiré lengths. The primitive reciprocal lattice vectors H H of a given harmonic (m, n) are given as the column vectors of Gmn = mG12 nG23. Inverting Gmn, we obtain the H 1 H T − moiré of moiré supercell in real space Amn = 2π (Gmn)− . The norm of the column vectors are the moiré of moiré H H lengths, denoted as λmn. For m = n = 1, the moiré of moiré length λ11 is explicitly given as

12 H (1 + δ)aM λ11 = , (S4) 2(1 + δ)(1 cos φ) + δ2 −

(12) 23 p 1 a1 a1 sin θ23 where φ = cos− (12) · (23) is the twist angle between the bilayer moiré supercells and δ = 1 is the a a sin θ12 | 1 || 1 | − 23 12 lattice mismatch between the two bilayer moiré supercells such that aM = (1 + δ)aM . Equation (S4) agrees with the first-order approximation for the moiré length for a twisted bilayer with a lattice mismatch [S3], with the lattice constant being the bilayer moiré length between L1 and L2. A dominant moiré of moiré length does not necessarily exist nor evolve smoothly under the continuous change of the twist angle. To see this, we will consider different harmonics of the higher-order moiré patterns. To find the H dominant harmonic for an arbitrary pair of twist angles, we calculate Amn for m , n 15 numerically and find the H | | | | ≤ (m, n) such that the norm of Gmn is the smallest, or, equivalently, that the moiré of moiré length λmn is largest. We are neglecting the cases where higher order harmonics dominate, such as the cases where θ12 and θ23 are different by more than a factor of 15. In those cases, the two bilayers moiré supercells have very different sizes and become essentially decoupled, which is not the focus of our study. Figure S2(a) shows the moiré of moiré harmonics for varying θ12 at a fixed θ23 = 2.8◦, indicating the non-smooth dependence of the dominant moiré of moiré length on the twist angle. In tTLG, there exists a supercell approximation when there is a clear dominant harmonic, that is when θ12 Nθ23 or θ23 Nθ12 for N Z. For example, at θ12 = 2.6◦, θ23 = 2.8◦, the dominant harmonic is (1, 1) [see ≈ ≈ ∈ Fig. S2(b)] and at θ12 = 1.35◦, θ23 = 2.8◦, the dominant harmonic is (2, 1) [see Fig. S2(d)]. However, there are cases where there is no clear dominant moiré of moiré. For example, in Fig. S2(c), it is difficult to visually discern a large repeating pattern and the estimated moiré of moiré lattice vectors fail to capture the relevant length scale. This is because near θ12 = 1.8◦, many harmonics, such as (3,2), (5,4), and (5,3), all have comparable lengths [see Fig. S2(a) the corresponding point].

II. MOMENTUM-SPACE MODEL

In this section, we oﬀer a detailed derivation of our momentum-space model and density of states formalism, and study the convergence as a function of the momentum space cutoﬀ radius. 3

H FIG. S2: (a) Moiré of moiré of lengths λmn as a function of θ12 for θ23 = 2.8◦. Each color corresponds to a different set of (m, n). The thick black line indicates the dominant length. (b)-(d) Examples moiré of moiré geometries, corresponding to the red crosses in (a). Top: red and blue scattered points are the lattice points of the bilayer moiré supercells between L1-L2 and L2-L3 respectively. Due to the different twist angles, the moiré lattice vectors are slightly rotated and have different lattice constants. Black vectors indicate estimated dominant moiré of moiré supercell lattice vectors. A blowup of the small boxed area is shown below, with points representing the atomic positions of each monolayer graphene, for L1-L2 on the left half and for L2-L3 on the right half. Red and blue vectors are the bilayer moiré lattice vectors of L1-L2 and L2-L3 respectively.

A. Detailed derivation of momentum-space model

To model the electronic structure of the tTLG system, we start from a tight-binding approximation for each individual layer; we take into account the interlayer hopping in a transverse tight-binding approximation between nearest neighbors. We start by writing the Hamiltonian for the trilayer as a sum of the following terms

3 H = H` + V `,`+1 + V `+1,` , (S5) l=1 l=1,2 X X where H` is the Hamiltonian for the `-th layer and V ij describes the interlayer hopping. For simplicity, we only consider the interlayer couplings between adjacent layers. DFT calculations predict that the interlayer coupling between L1 and L3 is roughly 10 times smaller than the coupling between adjacent layers [e.g., between L1 and L2] [S4]. In a second quantized notation, H` can be written as

` (`) (`) (`) (`) (`) (`) H = t c† (R )[c (R ) + c (R a ) + c (R a )] + h.c., (S6) − `,A `,B `,B − 1 `,B − 2 (`) RX (`) where c`,α† and c`,α are the creation and annihilation fermionic operators of the orbital α in layer l, a1,2 are the lattice vectors of layer l, and t is the hopping parameter between nearest neighbors. As for the interlayer coupling, we deﬁne the following overlap matrix element in the tight-binding basis

ij (i) (j) (i) (j) tαβ(R , R ) = i, R , α H j, R , β , (S7) D E where α and β denotes the sublattice degree of freedom. The interlayer Hamiltonian in the second-quantized notation is

ij (i) ij (i) (j) (j) V = ci,α† (R )tαβ(R , R )cj,β(R ). (S8) (i) (j) R ,α,XR ,β 4

We obtain the Hamiltonian in the momentum basis at a center site momentum k. Deﬁning Q(`) = k(`) + k for (`) (`) k Γ ∗. We perform the Fourier transform as follows ∈

(`) 1 (`) iQ(`) (R(`)+τ (`)) † α † c`,α(R ) = dk e · c`,k(`),α, Γ(`) Γ(`) | ∗| Z ∗ 1 (`) (`) (`) (`) p (`) iQ (R +τα ) c`,α(R ) = dk e− · c`,k(`),α, (S9) Γ(`) Γ(`) | ∗| Z ∗ p (`) (`) (`) (`) where the integral is over the Brillouin zone of the `-th layer, τA = 0, τB = 1/3(a1 + a2 ). The inverse of the transform in Eq. (S9) is

1 iQ(`) (R(`)+τ (`)) (`) c† = e− · α c† (R ), `,k(`),α (`) `,α Γ ∗ (`) | | RX p 1 iQ(`) (R(`)+τ (`)) (`) c (`) = e · α c`,α(R ), (S10) `,k ,α (`) Γ ∗ (`) | | RX (`) p where Γ ∗ is the area of the Brillouin zone in the `-th layer. The intralayer Hamiltonian in the Bloch basis can now | | be written as follows

(`) (`) (`) (`) ` t (`) (`) i(k k0 ) R ik s H = dk dk0 e − · e · i c† c (`) (`) `,k,A `,k0 ,B − Γ ∗ Γ(`) Γ(`) R(`) Z ∗ Z ∗ s(`) | | X Xi (`) iQ(`) s(`) i † (`) = t dk e · c`,k(`),Ac`,k ,B, (S11) − (`) Γ ∗ (`) Z Xs ik(`) R(`) (`) (`) where we use the Poisson summation formula, R(`) e · = Γ ∗ G(`) δk(`),G(`) . We also deﬁne si to describe | | (`) (`) (`) (`) the nearest neighbor separation between A andP B sublattices, whichP are given as s1 = 1/3(a1 + a2 ), s2 = (`) (`) (`) (`) (`) 1/3( 2a + a ), s = 1/3(a 2a ). The intralayer Hamiltonian in the basis of c (`) can then be written as − 1 2 3 1 − 2 `,k ,α (`) ` (`) 0 f`(Q ) H (Q ) = t (`) , (S12) − f ∗(Q ) 0 `

(`) iQ(`) s(`) where f`(Q ) = (`) e · i . The Hamiltonian is equivalent to the monolayer graphene tight-binding model at a si (`) (`) given momentum PQ [S5]. For the intralayer Hamiltonian, there is no constraint on Q .

† Similarly, we write the interlayer Hamiltonian in the c`,k(`),α basis

ij (i) (j) αβ (i) (j) † (j) V = dk dk ci,k(i),αTij (k , k )cj,k ,β, (S13) (i) (j) Γ ∗ Γ ∗ Z Z Xαβ where we use Eq. (S9) and

(j) ij (i) (j) 1 iQ(i) (R(i)+τ (i)) ij (i) (j) iQ(j) (R(j)+τ ) T (k , k ) = e · α t (R , R )e− · β . (S14) αβ (i) (j) αβ Γ ∗ Γ ∗ (i) (j) | || | R X,R p We now apply the two center approximation

(j) tij (R(i), R(j)) = tij (R(i) + τ (i) R(j) τ ), (S15) αβ αβ α − − β and write the interlayer coupling in terms of a two-dimensional Fourier Transform

tij (R(i), R(j)) = tij (R(i) + τ (i) R(j) τ (j)) αβ αβ α − − β dp ip (R(i)+τ (i) R(j) τ (j)) ij = e · α − − β t˜ (p). (S16) (2π)2 αβ Z 5

Plugging Eq. (S16) into Eq. (S14), the interlayer coupling matrix element in momentum space is

(j) ij (i) (j) 1 dp i(Q(i)+p) (R(i)+τ (i)) ij i(Q(j)+p) (R(j)+τ ) T (k , k ) = e · α t (p)e− · β αβ (i) (j) (2π)2 αβ Γ ∗ Γ ∗ (i) (j) | || | Z R X,R p dp (i) (i) ij iG(j) τ (i) (i) (j) iG τα ˜ β = Γ ∗ Γ ∗ e · tαβ(p)e− · δk+k(i) p,G(i) δk+k(j) p,G(j) | || | (2π)2 − − (i) (j) q G X,G Z 1 (i) (i) ij iG(j) τ (j) iG τα ˜ (i) (i) β = e · tαβ(k + k G )e− · δk(i) G(i),k(j) G(j) . (S17) Γ − − − (i) (j) | | G X,G (`) 2 1 In the last step, we use the Possion summation rule and Γ ∗ = 4π Γ − , where Γ is the monolayer unit cell area. | | | | | | We have obtained the scattering selection rule k(i) G(i) = k(j) G(j) for i = j 1, which imposes the constraint − − ± on the values of allowed k(`).

Combining the intralayer and interlayer terms, the Hamiltonian in the c† (`) basis can be represented as a 3 3 `,k ,α × block given in Eq. (1) of the main text.

B. Low-energy limit

We can greatly simplify the model by taking the low-energy limit. Each H` can be expanded around its Dirac point, k(`) = K + q(`), as a rotated Dirac Hamlitonian H` (q) for q = k + k(`) K : L` D − L`

iθ12 1 1 0 e q+ H (k) HD(q) = vF iθ , ≈ e− 12 q 0 − 0 q H2(k) H2 (q) = v + , ≈ D F q 0 − 0 e iθ23 q H3(k) H3 (q) = v − + , (S18) ≈ D F eiθ23 q 0 − (`) (`) where q = qx iqy. For the interlayer coupling, we substitute k = q + KL` into Eq. (S17), ± ±

ij (i) (j) 1 iG(i) τ (i) ij (i) (i) iG(j) τ (j) α ˜ α (i) (i) (j) (j) Tαβ(q , q ) = e · tαβ(k + KLi + q + G )e− · δq +KLi G ,q +KLj G . (S19) Γ − − (i) (j) | | G X,G For momenta near the Dirac point, since q(i) , k K , G(i) we can approximate t˜ij (k + K + q(i) + G(i)) | | | | | Li| | | αβ Li ≈ ˜ij (i) tαβ(KLi + G ). This approximation can lead to the suppression of particle-hole asymmetry in the tight-binding model [S6, S7]. Due to the rapid decay of the hopping parameter t˜(p) as p increases [S8], we keep only the ﬁrst shell in the summation in Eq. (S19):

3 ij (i) (j) ij T (q , q ) = T δ (i) (j) ij , αβ n,αβ q q , qn (S20) n=1 − − X ij ij 1 ij ij ij where q1 = KLi KLj, q2 = − (2π/3)q1 , and q3 = (2π/3)q1 (see Fig. S1b). We include out-of-plane relaxation ij −ij R ij ij R by letting tAA = tBB = ω0 = 0.07 eV and tAB = tBA = ω1 = 0.11 eV, which matches with the interlayer coupling in Nam and Koshino [S9] and Carr et al. [S6]. In matrix form,

ij ω0 ω1 ij ω0 ω1φ¯ ij ω0 ω1φ T1 = ,T2 = ,T3 = , (S21) ω1 ω0 ω1φ ω0 ω1φ¯ ω0 where φ = exp i 2π , φ¯ = exp i 2π . 3 − 3 6

FIG. S3: Top: DOS obtained using different sizes of momentum-space basis, nk. Bottom: errors in the DOS corresponding to the vertical lines on the top with the same color. The error is defined as the difference between the DOS value at the given energy and the DOS at the largest cutoff shown on the top. The dashed lines are guides to the eyes showing power law scaling of the error as a function of nk.

C. Density of States

The DOS at a given energy , (), for an incommensurate tight-binding model is deﬁned as [S10] D 1 N () = δ( ) ψ (r) 2, (S22) D N − n | n | r n=1 X X where the r sum is over all real space lattice positions, n is the band index, and ψn(r) is the corresponding eigenfunc- 2√ln 2 (x )2 tion. To obtain the DOS numerically, we use a Gaussian function φ (x) = exp 4 ln 2 −2 to approximate ,κ √πκ − κ the δ function, and κ is the full-width-half-maximum of the Gaussian, which determinesh the energyi resolution of the DOS [S11]. We can transform the DOS equation to momentum space:

2 () = N φ,κ(n,k) ψn,k dk, (S23) D 2 (`,`+1) | | Γ ∗ n α=XA,B `X=1,2 Z X where is a normalization constant, is an energy within the energy window [ ∆/2, + ∆/2), ∆ is the N n,k − energy interval, ψ and is an eigen-pair of the Hamiltonian (k) in Eq. (1) of the main text associated with n,k n,k H the center site k and band n. The integral is evaluated over the bilayer moiré Brillouin zone between layers ` and (`,`+1) (`,`+1) ` + 1, Γ ∗, and we discretize Γ ∗ using a 22 22 grid to evaluate the integral. We adapt κ based on the area (`,`+1) × of the integration domain Γ ∗ as θ`,`+1 changes. In order to make a direct comparison between the DOS at different twist angles, we need to properly normalize the DOS. For a given cutoff radius, we first calculate the DOS of the intralayer Hamiltonian only, which reduces to three independent copies of monolayer graphene. Near the charge-neutrality point, the DOS per eV per nm2 is given by [S5]

6 () = | 2| , (S24) D π vF where the prefactor includes a factor 3 from the number of layers as well as a factor of 4 from spin and valley degeneracies. We then obtain a normalization constant by ﬁxing the prefactor to the expected slope given in Eq. (S24) 7 and use the same constant for the DOS of the full Hamiltonian.

D. Convergence

The incommensurability of the tTLG system leads to an infinite number of coupled momenta within any finite cutoff radius. Due to the additional constraints we impose on the magnitude of G(`), we neglect degrees of freedom that can contribute to the low energy states. As a result, there is no guaranteed convergence. Figure S3(a)-(d) shows the DOS and the corresponding errors for different numbers of momentum degrees of freedom for tTLG with two different sets of twist angles. In both cases, as the cutoff increases, the error does not decay significantly. Note that in the case of θ12 = 2◦, θ23 = 2.4◦, the drop in error is most likely a numerical artifact and further increasing the cutoff will not likely to reduce the error. However, the physically relevant features, such as the magnitude of the DOS maximum and the positions of the VHS, are relatively stable as the cutoff increases. In contrast, Fig. S3(e), (f) shows the fast convergence of the DOS in tBLG as a function of cutoff radius. This is because in tBLG, increasing the cutoff radius does not change the number of relevant low-energy degrees of freedom. In this work, we choose a cutoff at the 4th honeycomb shell (i.e., k = 4 b(`) corresponding to 5 600 momenta). This choice was made by considering both c | | ∼ computational efficiency and the accuracy of physical properties of interest.

III. EFFECTIVE HAMILTONIAN AND RENORMALIZED FERMI VELOCITY

We examine the limit in which the momentum-space is truncated at the first honeycomb shell. The truncation gives rise to the following 14 14 Hamiltonian: × 1 12 12 HD(q + q1 )(T1 )† 1 12 12 HD(q + q2 )(T2 )†  1 12 12  HD(q + q3 )(T3 )† 12 12 12 2 23 23 23 (q) =  T1 T2 T3 HD(q) T1 T2 T3  . (S25) H  23 3 23   (T1 )† HD(q + q1 )   23 3 23   (T2 )† HD(q + q2 )   23 3 23   (T )† H (q + q )  3 D 3    This Hamiltonian acts on seven two-component spinors Ψ = (ψ11, ψ12, ψ13, ψ20, ψ31, ψ32, ψ33), where ` and j in ψ`j denote the layer and the momentum basis index respectively. Using this Hamiltonian, we can derive an expression for the renormalized Fermi velocity vF∗ . We first define the dimensionless quantities α12 = ω/vF kθ12 and α23 = ω/vF kθ23 , k = 8π sin (θij /2) ω = ω = ω H` where θij 3aG . For simplicity, we assume 0 1 and neglect the angular dependence in D by letting H` be an unrotated Dirac Hamiltonian: H` (q) = v σ q, where σ = (σ , σ ) is the Pauli matrix. The D D F · x − y zero-energy state of the Hamiltonian satisfies Ψ = 7 c Ψ = 0, where c is the column vectors of , and Ψ is H j=1 j j j H j the j-th component of the spinor Ψ. Therefore, we obtain the following relation between components of Ψ P ` 1 Ψ = (H )− T †ψ , (S26) j − D 20 where j = 4 (Ψ is not a state on L2 or ψ ). Using this, the effective Hamiltonian to the leading order in q is 6 j 20 3 2 vF 12 1 12 1 1 12 1 12 Ψ (q) Ψ = ψ† σ q + T (H (q + q ))− (σ q)(H (q + q ))− T † h | HD | i 1 + 6(α2 + α2 ) 20 · n D n · D n n 12 23 ( n=1 X h 23 3 23 1 3 23 1 23 + Tn (HD(q + qn ))− (σ q)(HD(q + qn ))− Tn † ψ20 · ) i = v∗ ψ† σ qψ , (S27) F 20 · 20 ? where the renormalized Fermi velocity vF is

2 2 vF∗ 1 3(α12 + α23) = − 2 2 . (S28) vF 1 + 6(α12 + α23) 8

Figure S4(a) shows the vF∗ to vF ratio as a function of θ12 at a few values of θ23. As θ23 increases, the vF∗ /vF ratio approaches the tBLG curve. Figure S4(b) shows vF∗ /vF for equal twist angles, which shows that perturbation theory predicts that vF∗ can still go to zero at 1.72◦. However, in our numerical calculation using the full Hamiltonian, we do not observe a complete ﬂattening of bands at this twist angle.

FIG. S4: (a) The ratio of renormalized Fermi velocity vF∗ to the monolayer Fermi velocity vF as a function of θ12 for given values of θ23. Black dashed line shows the tBLG vF∗ /vF ratio. (b) vF∗ /vF ratio as a function of twist angle for θ12 = θ23.

FIG. S5: Comparison between the renormalized Fermi velocity vF∗ of the Hamiltonian in Eq. (S25) calculated analytically (solid line) and numerically (dashed line).

Finally, we show that our assumption in the analytic calculation of an unrotated Dirac Hamiltonian for the intralayer Hamiltonian and ω0 = ω1 does not significantly change the magic angle estimate. Figure S5 compares the vF∗ obtained analytically and numerically and show that the two curves and the magic angle do not differ significantly. In the numerical calculation, we diagonalize the 14 14 Hamiltonian with rotated Dirac equation for the intralayer terms × and ω0 = 0.07 eV, ω1 = 0.11 eV for the interlayer terms. At θ23 = 2.5◦, the magic angle obtained analytically and numerically differ by 1.1%.

IV. COMPARISON TO OTHER MODELS

In this section, we compare our results to two other works [S1, S2]. We first compare our results with the model proposed by Mora et al.[S1] and use it to gain further insights into our findings. In this alternate model, a different 9

FIG. S6: Comparison of the bilayer moiré Brillouin zone geometry between our model and Mora et al. [S1] model with θ12 = θ23. Left: two bilayer moiré Brillouin zones are misaligned by a small twist angle; right: the two bilayer moiré Brillouin zones are approximated to be aligned.

FIG. S7: The momentum degrees of freedom in the low-energy limit of the simplified model for (a) θ12 = θ23 and (b) 2θ12 = θ23 with high symmetry points. momentum-space basis is used by aligning the two bilayer moiré Brillouin zones [Fig. S6]. This approximation ignores the incommensurability of the system, making a two-dimensional momentum space crystal with the periodicity of the bilayer moiré Brillouin zone. As a result, the problem’s complexity reduces to that of a bilayer. Formally, the Hamiltonian can still be written as the 3 3 block as in Eq. (1) in the main text, but the size of the basis is reduced × to be on the same order as tBLG. We implemented two cases: (1) θ12 = θ23 and (2) 2θ12 = θ23. Figure S7 shows the momentum-space basis for these two cases. In case (2), the larger bilayer Brillouin zone (L1-L2) is folded onto the smaller Brillouin zone (L2-L3) in momentum space. This model essentially describes a system consisted of 2 2 × L1-L2 moiré supercell and a L2-L3 moiré supercell. Figure S9 shows a comparison between the DOS obtained from the two models. We keep the values of ω0 and ω1 the same as our model and use the same approach to normalize the DOS for a direct comparison. We cut off the basis at the 4th shell and use a grid size 22 22 for the density of states. × The Gaussian FWHM we use is 5 meV for θ < 2◦ and 8 meV for θ 2◦, where θ is the twist angle that determines ≥ the size of the Brillouin zone. For θ12 = θ23, Fig. S8(a) shows the DOS obtained with the simplified model, which agrees qualitatively with the DOS from our model [Fig. 3(a) of the main text]. However, here the DOS has the sharpest peak between 1.7◦ and 1.8◦, and at 2.1◦ the VHS have a larger width compared to our model. Figure S8(b)-(d) shows that the location of peaks away from the CNP are also very different from our model. For 2θ12 = θ23, the two models predict similar trend for the VHS evolution, and the simplified model makes the right prediction for the magic angle. This is expected from perturbation theory, since the magic angle condition does 10

FIG. S8: DOS obtained with the Mora et al. [S1] model DOS on a logarithmic color scale at equal twist angles (same color scale as Fig. 3(a) in the main text for a direct comparison) (b)-(d) DOS states along the black dashed line in (a) for (b) θ = 1.8◦, (c) θ = 2.1◦, (d) θ = 2.5◦, where the black solid lines are obtained using the Mora et al. [S1] model, and the blue dashed lines are obtained using our full model. not rely on the existence of a moiré of moiré cell [as was shown in SectionIII]. However, the magnitude of the DOS differs significantly between the two models. This is because there are two flat bands near the CNP in the simplified model, whereas in our full model, there is a large number of nearly overlapping flat bands due to incommensurability [Fig. S10]. Figure S10 compares the band structure from our model and the simplified model. The two band structures are qualitatively similar but our model shows a large number of bands due to the lack of a periodic Brillouin zone. Furthermore, the aligned-bilayer approximation will exclude correlated phases that depend on band-hybridization or symmetries from the moiré of moiré length scale. Note that we do not plot the relative layer weights (color) of the band structure in the simplified model because of the way that the Brillouin zone is wrapped – the L1 degrees of freedom are wrapped on top of the L3 degrees of freedom. Therefore, the wavefunction weights from the two models are not directly comparable for this particular high symmetry line cut.

FIG. S9: DOS as a function of θ12 with 2θ12 = θ23 using (a) our full model and (b) the Mora et al. [S1] model, both on a logarithmic color scale.

We can use these results to further support our argument of bilayer moiré hybridization at equal twist angles. In 11

FIG. S10: Comparison of band structures and DOS at θ12 = 1.4◦, θ23 = 2.8◦ between our model (top) along the green dashed line in Fig. S1(b) and the Mora et al. [S1] model (bottom) along the black dashed line in Fig. S7(b). In (a), colors represent the weight of the wavefunctions at the center site. Red, blue, and green represents weights purely from L1, L2, L3 respectively, and colors in between represent hybridization between diﬀerent layers. A colormap is provided on the top left corner. The DOS from the two models are shown on the same scale.

this simplified model, sharpest VHS occur between 1.7◦ and 1.8◦, which is in better agreement with the magic angle prediction from perturbation theory. In our model, the sharpest peak and the narrowest width occurs at a larger angle (2.1◦). If this phenomenon is caused by moiré hybridization, the simplified model would not have it since it does not have the moiré of moiré scale. Indeed, the DOS from the two models differ most significantly at 2.1◦ [see Fig. S8(c)]. As we argue in the main text, adding electrons from the CNP at a low carrier concentration on the order of the tTLG moiré of moiré cells fills one flat band near the CNP in Fig. S10(a) at a time. Injecting electrons at a carrier concentration comparable to the bilayer moiré cell density would fill all these flat bands near the CNP. The simplified model can again be used to understand this argument. The model also predicts some band flattening at certain twist angles, but there are only two flat bands near the CNP [Fig. S10(c)]. Filling electrons to these two bands is equivalent to filling the bilayer moiré cell, since their momentum-space basis has the periodicity of bilayer moiré Brillouin zone and there is no moiré of moiré length in this model. These two flat bands near the CNP can be qualitatively considered as the limit where all the flat bands from our model overlap exactly on top of each other. Therefore, in terms of filling the supercell, filling the two flat bands from the simplified model is equivalent to filling all flat bands in the full model. In addition to its inability to make predictions about electronic behaviors at the moiré of moiré scale, another major limitation of the model is its difficulty to generalize to arbitrary twist angles. For each set of twist angles on a different (m, n) harmonic, it requires the derivation of a new basis by folding the bilayer moiré Brillouin zone, while our model’s basis is insensitive to the choice of angles and overcomes this limitation. We can also use our model to study the case where L1 and L3 are twisted in the same direction (when θ12 and θ23 take opposite signs). This case has been studied theoretically by Amorim and Castro [S2] and its spectral properties have been investigated experimentally by Zuo et al. [S12]. Unlike our model, Amorim and Castro [S2] does not 12 take the low-energy limit [see SectionIIB]. Figure S11 shows the band structure and the corresponding DOS of θ = 2.81◦, θ = 2.1◦ obtained with our model, which is the same case as Figs. 1(a) and 2 presented in Amorim 12 − 23 and Castro [S2]. The results from the two models show an agreement, with the same VHS positions. The difference in the band structure can be most likely attributed to the different ways of truncating the momentum-space bases between the two models.

FIG. S11: (a) Band structure along a high symmetry line that connects the Dirac points of the three layers and (b) density of states at θ12 = 2.81◦, θ23 = 2.1◦. The colormap in (a) is the same as in Fig. S10(a). −

[S1] Christophe Mora, Nicolas Regnault, and B. Andrei Bernevig, “Flatbands and Perfect Metal in Trilayer Moiré Graphene,” Phys. Rev. Lett. 123, 026402 (2019). [S2] B. Amorim and Eduardo V. Castro, “Electronic spectral properties of incommensurate twisted trilayer graphene,” arXiv e-prints , arXiv:1807.11909 (2018), arXiv:1807.11909 [cond-mat.mes-hall]. [S3] Matthew Yankowitz, Jiamin Xue, Daniel Cormode, Javier D. Sanchez-Yamagishi, K. Watanabe, T. Taniguchi, Pablo Jarillo-Herrero, Philippe Jacquod, and Brian J. Leroy, “Emergence of superlattice Dirac points in graphene on hexagonal boron nitride,” Nature Physics 8, 382–386 (2012). [S4] Stephen Carr, Chenyuan Li, Ziyan Zhu, Efthimios Kaxiras, Subir Sachdev, and Alexander Kruchkov, “Ultraheavy and ultrarelativistic dirac quasiparticles in sandwiched graphenes,” Nano Letters 20, 3030–3038 (2020). [S5] A. H. Castro Neto, F. Guinea, N. M. R. Peres, K. S. Novoselov, and A. K. Geim, “The electronic properties of graphene,” Reviews of Modern Physics 81, 109–162 (2009). [S6] Stephen Carr, Shiang Fang, Ziyan Zhu, and Efthimios Kaxiras, “Exact continuum model for low-energy electronic states of twisted bilayer graphene,” Physical Review Research 1, 013001 (2019). [S7] Shiang Fang, Stephen Carr, Ziyan Zhu, Daniel Massatt, and Efthimios Kaxiras, “Angle-Dependent Ab initio Low- Energy Hamiltonians for a Relaxed Twisted Bilayer Graphene Heterostructure,” arXiv e-prints , arXiv:1908.00058 (2019), arXiv:1908.00058 [cond-mat.mes-hall]. [S8] Raﬁ Bistritzer and Allan H. MacDonald, “Moiré bands in twisted double-layer graphene,” Proceedings of the National Academy of Science 108, 12233–12237 (2011). [S9] Nguyen N. T. Nam and Mikito Koshino, “Lattice relaxation and energy band modulation in twisted bilayer graphene,” Phys. Rev. B 96, 075311 (2017). [S10] Stephen Carr, Daniel Massatt, Shiang Fang, Paul Cazeaux, Mitchell Luskin, and Efthimios Kaxiras, “Twistronics: Manipulating the electronic properties of two-dimensional layered structures through their twist angle,” Phys. Rev. B 95, 075420 (2017). [S11] Daniel Massatt, Stephen Carr, Mitchell Luskin, and Christoph Ortner, “Incommensurate heterostructures in momentum space,” Multiscale Modeling & Simulation 16, 429–451 (2018). [S12] Wei-Jie Zuo, Jia-Bin Qiao, Dong-Lin Ma, Long-Jing Yin, Gan Sun, Jun-Yang Zhang, Li-Yang Guan, and Lin He, “Scanning tunneling microscopy and spectroscopy of twisted trilayer graphene,” Physical Review B 97, 035440 (2018).