Deep-Learning Database of Density Functional Theory Hamiltonians for Twisted

Materials

Ting Bao,1, 7 Runzhang Xu,1, 7, ∗ He Li,1, 2 Xiaoxun Gong,3 Zechen

Tang,1 Jingheng Fu,1 Wenhui Duan,1, 2, 4, 5, † and Yong Xu1, 4, 6, ‡
1
State Key Laboratory of Low Dimensional Quantum Physics and
Department of Physics, Tsinghua University, Beijing, 100084, China
2
Institute for Advanced Study, Tsinghua University, Beijing 100084, China
3
School of Physics, Peking University, Beijing 100871, China
4
Frontier Science Center for Quantum Information, Beijing, China
5
Beijing Academy of Quantum Information Sciences, Beijing 100193, China
arXiv:2404.06449v1 [cond-mat.mtrl-sci] 9 Apr 2024

6
RIKEN Center for Emergent Matter Science (CEMS), Wako, Saitama 351-0198, Japan
7
These authors contributed equally
Moiré-twisted materials have garnered significant research interest due to their distinctive prop-
erties and intriguing physics. However, conducting first-principles studies on such materials faces
challenges, notably the formidable computational cost associated with simulating ultra-large twisted
structures. This obstacle impedes the construction of a twisted materials database crucial for data-
driven materials discovery. Here, by using high-throughput calculations and state-of-the-art neural
network methods, we construct a Deep-learning Database of density functional theory (DFT) Hamil-
tonians for Twisted materials named DDHT. The DDHT database comprises trained neural-network
models of over a hundred homo-bilayer and hetero-bilayer moiré-twisted materials. These models
enable accurate prediction of the DFT Hamiltonian for these materials across arbitrary twist angles,
with an averaged mean absolute error of approximately 1.0 meV or lower. The database facilitates
the exploration of flat bands and correlated materials platforms within ultra-large twisted structures.

INTRODUCTION ods, such as DFT, tight-binding method, and continuum

model, are employed to make predictions for guiding ex-
periments. However, for twisted materials, all these the-
The discovery of correlated states in magic-angle
oretical methods encounter the vital efficiency-accuracy
twisted bilayer graphene (TBG) [1, 2] has sparked great
dilemma. For instance, DFT methods scales as O(N 3 )
research interest in the emergent physics and exotic prop-
with N being number of atoms per periodic unit cell,
erties in moiré-twisted two-dimensional (2D) materials.
which limits the sizes of twisted structures to be stud-
In moiré twist systems, the relative twist rotation be-
ied. The empirical tight-binding methods or continuum
tween the two constituent van der Waals (vdW) layers
models are much more efficient, but suffers from the crit-
forms moiré patterns with tunable periodicity, generating
ical problem of low accuracy. The revolutionary success
a moiré potential field that strongly influences the elec-
of artificial intelligence, especially deep learning neural
tronic structures and introduces exotic quantum effects
networks, in many fields [18–20] brings solution to this
such as correlated states, band topology, spin textures,
dilemma. In previous work DeepH [21], Li et.al. show
and ferroelectric polarization [3]. These twist-induced in-
that a neural network (NN) model trained on small struc-
teresting underlying physics have been intensively stud-
tures can be generalized to tackle the DFT Hamiltoni-
ied on TBG, twisted bilayer CrI3 [4, 5], twisted bilayer BN
ans of large twisted structures at significantly reduced
(TBBN)[6, 7], twisted transition metal dichalcogenides
computational cost. However, only TBG, twisted bilayer
(TTMDs) [8] and so on. Other twist-induced properties,
bismuthene and twisted MoS2 are studied using DeepH,
like moiré excitons [9–12], Wigner crystal states [13], and
which is not enough to validate the effectiveness. A more
spin- and bond-density waves [14, 15] further expand the
systematic learning on DeepH methods is in urgent need,
knowledge of twisted materials. These intriguing effects
which is not only to check the validness and universality
and physics, coupled with the rich external modulations,
of DeepH but also to meet the demand of getting the
position twisted materials not only as ideal quantum
DFT-level Hamiltonians for more materials and boost
simulators [16] in condensed-matter research but also as
further research.
practical platforms for twistronic applications.
Despite the growing interest in twisted materials and Here, we report the construction of a deep-learning
the related physics, the field faces fundamental challenges database of DFT Hamiltonians for twisted materials
from both experimental and theoretical perspectives [17]. named DDHT as shown in Fig. 1. DDHT currently cov-
In experiments, the synthesis of high-quality twisted sam- ers DeepH NN models for 124 types of 2D homo-bilayer
ples and the precise control of twist angles are rather and 5 types of hetero-bilayer twist materials and can
challenging. From theoretical and calculation perspec- efficiently predict accurate DFT Hamiltonians of their
tive, various kinds of numerical and theoretical meth- twisted structures across arbitrary twist angles. For each
 2

DFT calculations

Crystal graph DeepH Hamiltonian

DeepH code
Twist angles
Perturbations & Slidings
Training
Training dataset

Diversity
Constituent
materials
• Band structures

• Band topology

• Projected density of
DeepH model states
Moiré-twisted C2DB
database
structures • ……

FIG. 1. Schematic diagram of construction and application of DDHT database. Uniformly distributed interlayer slidings
together with random atomic perturbations are introduced in the non-twisted bilayer supercells to construct training datasets.
Graph neural networks are employed here, using materials structures as input and their DFT Hamiltonians as output. Deep
neural network models are trained by DFT calculation results of training structures, and then generalized to study diverse
moiré-twisted structures with varying twist angles and constituent materials selected from C2DB. From the DFT Hamiltonians
predicted by DeepH, many physical properties can be derived, including band structure, band topology, projected density of
states, etc.

type of material, the NN model is trained by the state- celerate the exploration in this field.
of-the-art deep-learning representation of DFT Hamilto- It is worth mentioning that unlike traditional material
nians implemented in the E3-equivariant DeepH method databases like Materials Project [23], ICSD [24] or C2DB
[21, 22]. We demonstrate the high reliability and DFT- [25, 26], which includes specific materials corresponding
comparable accuracy of all NN models in our database to discrete points in the materials parameter space, our
by the very low validation loss achieved in the NN train- DDHT is applicable to twisted structures at arbitrary
ing process and using the DFT-calculated Hamiltonians twist angles and interlayer fluctuation of provided types
at selected large twist angles as benchmark. Further- of material, covering wide areas in the materials parame-
more, taking 2D black phosphorene (bP) as an exhibi- ter space with infinite single structure cases inside. Fur-
tion, we use the corresponding NN model in DDHT to thermore, considering each model in the DDHT as an
obtain electronic band structure and find the ultra-flat expert and adding a simple classifier as the gate on types
valence bands in twisted bilayer bP down to twist angles of material, out DDHT itself is actually serving as a mix-
inaccessed by conventional DFT methods, showing the ture of experts (MoE) model [27] with each NN item ex-
effectiveness and efficiency of our database in getting ex- perienced in one type of twisted material. Thus DDHT
otic electronic properties of ultra-large twisted materials. has dualism of database and MoE model, we suppose
In our exhibition of twisted bP and probably many other this dualism is a new formunism and will be a necessity
twisted materials systems, the flat bands usually have of future databases in the AI era.
bandwidth droping to only a few meV at small-angle limit
(<2.0 ◦ ) and their real space projection can be highly
localized, giving rise to rich exotic properties from dop- RESULTS
ing modulation and strong correlation. DDHT not only
provides an optimal parameter strategy for NN training
Construction of Database
and strongly supports the universality of DeepH method,
but also offers prior DFT-level knowledge of Hamiltoni-
ans and electronic structures for researchers in the field Candidate 2D materials considered in the current
of twisted materials and twistronics and thus greatly ac- database are adopted from C2DB [25, 26]. They are fur-
ther screened to construct twisted materials structures
 3
Colorful-0129 换个fancy 的logo？

Candidate materials Training data Neural network

selection generation training and validation

C2DB Relaxed bilayer candidates Processed dataset

DDHT

Structure relaxation by VASP

DFT Hamiltonians extraction

3x3 supercell
Stable Training DeepH-E3
NN on given datasets

Archive in database
Non-magnetic

Uniform distributed
interlayer shift
MSE < 1e-6
Atom ≤ 8 Learning rate < 3e-5
Element ≤ 3 Atomic perturbation

Trained Models
Unrelaxed bilayer Training structures
candidate material

DFT calculation Performance check on

Human-picked candidates using OpenMX benchmark structures

FIG. 2. Workflow of constructing deep-learning database. All the steps are classified into three parts: candidate selection,
dataset generation, and neural network training and validation. The parameters used in construction are explicitly displayed
by typical values for key steps. The eventually obtained DDHT database can be further used for unlimited cases.

and training datasets (left part of Fig. 2). In total, 124 als host rich structural diversity in terms of twist angle
types of 2D materials (Table S1) in the C2DB are se- and constituent materials, which should be captured by
lected following three criteria: non-magnetic, thermally atomic structures in the training datasets for accurate
and dynamically stable, structurally simple (≤ 8 atoms R 7→ HDeepH-E3 mapping by the trained models. This
per unit cell, ≤ 3 elements). Their bilayer unit cells are requirement is satisfied by preparing atomic structures
constructed and fully relaxed by DFT calculations using in datasets respectively (or separately) for each type of
Vienna ab initio simulation package (VASP) [28, 29] (see twisted materials, and introducing diverse enough vari-
Note 1 in Supporting Information for details of screening ations to local atomic positions and interlayer stackings
and relaxation). This allows us to obtain the optimal to cover all possible local environments in twisted bi-
structural parameters, including lattice constants, stack- layer structures. For instance, TBG and TBBN have
ing orders, and interlayer spacings. These structural pa- different constituent 2D materials (graphene and BN),
rameters will be used to construct the non-twisted super- and the non-twisted supercells in their datasets are gen-
cell structures in the datasets and the corresponding 124 erated using graphene and BN bilayer unit cells respec-
(5) types of to-be-predicted moiré-twisted homo-bilayer tively, while the twisted heterobilayer graphene/BN have
(hetero-bilayer) structures with varying twist angles θs. heterogeneity in layers and thus uses graphene/BN unit
These candidates not only cover most of the intensively cells for generating the dataset. The relative rotations
studied 2D materials and their twisted structures, such as induced by twisting between the two vdW layers can be
graphene, black phosphorene (bP), and transition metal treated as locally approximated interlayer shear slidings
dichalcogenides (TMDCs), but also include most light combined with perturbations to atomic positions in non-
elements in the periodic table (Fig. S1). twisted bilayer structures. Therefore, for each type of
The schematic diagram of constructing and using our twisted material, the atomic structures in its dataset are
database is demonstrated in Fig. 1 using TBG as an prepared (”Training data generation” part in Fig. 2)
example material. Here, the construction involves both by 1) making non-twisted supercells with sufficient large
the preparation of appropriate datasets and the train- lateral size from the most optimal bilayer unit cells to
ing of deep neural network models. The usage refers include all neighbors within cut-off radius Rc of local
to the prediction of DFT Hamiltonians for twisted ma- atomic basis and also to avoid interactions from periodic
terials with any given twist angle as well as various images, and 2) introducing uniform distributed interlayer
kinds of derived physical properties. We first discuss the slidings between the two vdW layers and random pertur-
preparation of training datasets. Moiré-twisted materi- bations to the atom positions in these non-twisted su-
 4

percells to cover the local structural diversity in twisted efficiency-limiting DFT calculations in computational re-
structures, with the mesh density of slidings and the up- search flow in twisted materials and still output Hamil-
per limit of perturbations determined based on balanced tonians with similar accuracy.
accuracy and efficiency. The local-atomic-basis Hamil-
tonians HDFT of all these non-twisted, perturbed, and
shifted supercell structures in the dataset are calculated Accuracy of the Database
self-consistently by first-principles DFT methods and fur-
ther processed to extract the decomposed hopping matrix To assess the accuracy and reliability of these DeepH-
segments Hij that are in one-to-one correspondence with E3 models in our DDHT database in predicting Hamil-
local atomic structures Rij within Rc . After represent- tonians and electronic structures of twisted materials, we
ing Rij by crystal graphs [30], we finally prepared the perform comprehensive statistics over mean absolute er-
complete training dataset for each twisted material by rors (MAEs) between the DeepH-E3-predicted and the
incorporating these Hij –Rij pairs. DFT-calculated Hamiltonians for selected rigid-twisted
Based on each training dataset, an E(3)-equivariant structures of all 124 homo-bilayer materials. Considering
deep neural network model is built and trained (”Neural the heavy computational demand of DFT Hamiltonians
network training and validation” part in Fig. 2) using of twisted materials, we only cover the twisted supercells
the state-of-the-art DeepH-E3 code [21, 22]. To ensure with either the largest two commensurate twist angles
the high accuracy and generalizability of the model, the (e.g. 21.79◦ and 13.17◦ corresponding to twist index 2-1
training process runs for a sufficient number of epochs and 3-2 for hexagonal lattice) or the number of atoms
and stops when the validation loss reaches a strict cri- less than 120 in the MAE statistics. For each type of
terion (see Sec. METHODS for details). Based on a material, the absolute-error matrices are calculated by
self-made high-throughput computation framework, the taking the absolute difference between respective hop-
dataset preparation and the model training for all 124+5 pings of the predicted Hamiltonians of its twisted bilayer
types of twisted materials are carried out in high effi- structures and that of the DFT calculated ones. These
ciency, which in the end output 124 + 5 E(3)-equivariant absolute error values are further averaged with respect
deep neural network models and their corresponding to hopping pairs and containing chemical elements, re-
non-twisted bilayer unit cells that both constitute the sulting in the material-wise MAEs as shown in Fig. 3.
database. The DFT calculations in preparation of train- The overwhelming majority of hopping MAEs are in the
ing datasets and the process of model training are both, sub-meV order of magnitude, which are comparable to
to some extend, computationally expansive. Typically, or even less than commonly recognized errors in gen-
for each type of material, they will take 1-3 days on eral DFT calculations. It’s safe to say 1) the DeepH-E3
a 64-core CPU node and 2-7 days on a NVIDIA V100 method can be efficiently generalized to much more ma-
GPU card. However, the resulting well-trained DeepH- terials using the workflow and parameter strategy here
E3 models in our database have strong generalizability in the DDHT 2) Hamiltonian predictions by models in
and universality over a wide range of θs (from 21.79◦ all our DDHT database are indeed at DFT accuracy. By
the way down to less than 1◦ ) and are capable of predict- expanding the MAE heatmaps with respect to orbitals
ing DFT-accuracy HDeepH-E3 of ultra-large (ultra small- of different element pairs (see Note 5 in Supporting In-
angle) twisted structures beyond the reach of common formation), we find no significant large MAEs between
DFT codes with an acceptable cost. For example, the these explicitly presented orbital-wise hoppings which
computational time of predicting HDeepH-E3 scales only may qualitatively weaken the above conclusion.
linearly with the number of atoms in lattice cells, allow- It’s also important to note that the models in our
ing for the calculation of DFT-accuracy Hamiltonian for database work well for twisted materials both with and
magic-angle TBG in just about 100 seconds [21]. without spin-orbit coupling (SOC) in our demonstrations
Other than the direct usage of our database in pre- as long as the datasets (thus the DeepH-E3 models) in-
dicting HDeepH-E3 by direct neural-network mapping for corporate SOC for respective (SOC-embedded) materi-
twisted structures with arbitrary θs and situation of als.
interlayer fluctuation, several important physical prop- We further verify the accuracy and effectiveness of the
erties can be further calculated from the predicted models in our database by calculating the electronic band
HDeepH-E3 sparse matrices (dashed box on the right in structures from the above predicted Hamiltonian matri-
Fig. 1), which may include the eigen values and states, ces and presenting the close fit between the predicted
band structures, total and projected density of states energy bands (energy dispersion relations) and the DFT-
(DOS), and possible correlated flat bands near the Fermi calculated ones. Here, two representative materials, bP
level and their real-space projections. It should be em- and 2H-TaS2 , are chosen as demonstration based on the
phasized that our database, serving as a MoE for DFT general research interest in their unique features, namely,
Hamiltonian prediction, offers an ultra-efficient way (en- the potential flat bands in twisted rectangular buckled
hanced by orders of magnitude [21]) to bypass the most lattice [15] of semiconducting bP and the interplay among
 5

a Elementary substance b Ternary compound

1.0 1.0
MAE (meV)

MAE (meV)
0.5 0.5

0.0 Bi2 C2 P4 Sn2

0.0 r2 c c2 N2 Te 2 S4 e4 e c r e2 e
BiIS Br 2N 2Z CCl 2S 2Cl 2H 2S Cl 2Hf 2 CS 2Ta 2 a 2Ge 2 Hf 2I 2N 2 I 2N 2Zr 2 MgAl 2 gAl 2S MoSS NCl 2S 2 PTe 2Z 2 eSb 2S TiSeT ZnIn 2S 4
G M T

c Binary compound
1.0 1.0
AB type A2B3 type
MAE (meV)

MAE (meV)
0.5 0.5

0.0 0.0
BN Zr 2 Cl 2Zr 2a Cl 2Zr 2b Cu 2I 2 Ga 2
S2
Ga 2
S 2 a 2Se 2 a 2Te 2 GeS GeS
e
GeT
e
In 2S 2 In 2S
e 2 i Te 2 SnSe
V 2S 2
Bi2Se3 Bi2Te3 In2Se3 Sb2Te3 W2N3
Br 2 G G N2

1.0
AB2 type
MAE (meV)

0.5

0.0 I2 l 2 aI 2 l2 I2 I 2 O 2 i 2a i 2b l2 I2 O2 S2 e 2 e 2 iO 2 iS 2 Se 2 Te 2 r 2a r 2b Cl 2 I 2a I 2b
Br Ba aBr aC Br dC Cd CrS rSe Ge Br fCl HfS fSe fTe gBr gC Mg
2 2 2 2 2 2 2 2 2 2 2 O2 O2
Ba C C C
Cd C C Ge HfB HfB Hf H H H M M M
o Mo oS MoT N N Ni Ni PbB bB Pb Pb Pb Pb Pd
M P

1.0
AB2 type
MAE (meV)

0.5

0.0 a b a b r 2 Cl 2 I 2a a a
Se dTe PtO PtS tSe tTe bSe nI 2 nI 2 nS 2 nS 2 nSe rB
2 2 2 2 2 2 2 2
I 2b 2
b e2 e 2 Br 2 iCl 2 Ti
I2
Te WO WS
2 2 2
Se
2
Te nBr nC
2 2 l 2 Br 2 Cl 2 I2 2 b
Zr ZrS Se 2 Te 2
Sr Sr TaS TaS TaS TaT Ti
2
Pd P P P S S S S S S S Sr T Ti W W Z Z Zr Zr Zr Zr

d Hetero structures
1.0
AB3 type
MAE (meV)

MAE (meV)

1.0
0.5
0.5

0.0 Al2Cl6 Bi2I6 In2Br6 Ir2Br6 Ir2Cl6 Rh2Br6 Sc2Cl6 Y2I6

0.0 Bi2Se3 Graphene MoS2 MoTe2 WS2
Bi2Te3 BN MoSe2 WSe2 WSe2

FIG. 3. The statistics over MAEs between the DeepH-E3-predicted HDeepH-E3 and the DFT-calculated HDFT of 124(5) types of
twisted homobilayer(heterobilayer) materials covered in our database. (a-d) shows the MAEs of elementary substances, ternary
compounds, binary compounds (divided into four prototypes AB, AB2 , A2 B3 and AB3 ) and hetero structures, respectively, by
bar plot. The height of each bar displays the MAE for one specific 2D material, which is obtained by averaging Hamiltonian
errors over all atoms and orbitals for its twisted structures with the largest two commensurate twist angles θ = 21.79◦ (2-1) and
13.17◦ (3-2), except for θ = 10.87◦ (7-4) and 8.49◦ (9-5) in rectangular P4 (bP) and θ = 10.87◦ (7-4) and 8.49◦ (9-5) in square
SnS2 . The predicted MAEs for most materials are in the order of sub meV, with only the twisted heterobilayer Bi2 Se3 /Bi2 Te3
and WS2 /WSe2 being exceptional (about 1.2 meV).

twisting, SOC, charge density wave, and superconduc- observed in all other twisted homobilayer materials cov-
tivity in metallic TaS2 [31]. At large twist angles, the ered in our database, as shown in the left panels of Fig.
energy bands obtained from the DeepH-E3 predictions SX and also reported in previous studies [21, 22].
(blue dots) and the DFT calculations (red curves) are
almost identical (left panels in Fig. 4a, b, d, e) in the
0130! The twisted heterobilayers induce extra complex-
ity in crystal structures and interatomic hopping rela-
energy range near the Fermi level. Similar results are tions, which is more challenging for DeepH-E3 mod-
 6

els to give accurate predictions. We test the valid- domain beyond the computational reach of most DFT
ity regarding the heterogeneity in their two constituent codes while maintaining acceptable computational costs.
layers in several selected twisted heterobilayer materi- To illustrate this capability, we have chosen twisted bP
als, such as Bi2 Se3 /Bi2 Te3 , graphene/BN, MoS2 /MoSe2 , bilayer structures at various twist angles as exemplars.
WS2 /WSe2 , and MoTe2 /WSe2 , and include the DeepH- As shown in the Fig. 4a, as the twist angle of bP ap-
E3 models of these materials in our database. Our proaches smaller values, the time required for the NN
database outputs accurate predictions for all these model to perform Hamiltonian inference, which is di-
twisted heterobilayers considered (Fig. 4c, f and Fig. rectly proportional to the number of atoms in the unit
SXXX-SXXX) and therefore holds the accuracy and ef- cell, increases rapidly but remains within the order of
fectiveness toward heterogeneity in constituent layers. minutes. It is noteworthy that the widths of energy band
Take twisted strong-SOC Bi2 Se3 /Bi2 Te3 with θ being near the Fermi energy diminish as the twist angle de-
21.79◦ and 13.17◦ as an example, the predicted and DFT- creases, as shown in Fig. 4b, revealing the band width
calculated energy bands have negligible difference (with of the highest valence band and the average band width
MAE in the order of XXX meV), and more importantly, of the top 10 valence bands with respect to the twist
the non-trivial Z2 topology of both twisted structures is angle of bP. The top two valence bands in Fig. 4a, d
also well reproduced by prediction. Therefore, as a step narrows their band width along the X-Γ high-symmetry
forward, our database firstly generalized the DeepH-E3 path with θ decreased by about 1.93◦ , and they will be
method and models to twisted heterobilayers with com- potentially flattened throughout the whole momentum
parable accuracy to homobilayer cases. space in twisted supercells at smaller θs. In this study,
In addition to the DFT band structures, the predicted we examine two bP twisted supercells with θ = 2.36◦
Hamiltonians by our database can be used to calculate (4756 atoms in unit cell) and 1.59◦ (10588 atoms in unit
the electron density of states (DOS) through identify- cell), which are relaxed by a machine-learning force-field
ing the basis projections of their eigenvectors. The right model (GAP+R6) for phosphorene [32] to eliminate the
panels in Fig. 4a-f display the total (grey curves) and unreasonable band crossings found in ion-clamped struc-
the projected DOS (colored curves) of twisted bP, 2H- tures at small θs. Obtaining the band structures of the
TaS2 , and Bi2 Se3 /Bi2 Te3 , with DOS of all other large- two bP twisted supercells with thousands of atoms using
angle twisted structures presented in the right panels of ab initio DFT methods consumes unacceptable compu-
Fig. SX2. The predicted total and projected DOS pro- tational resources, however, they can be obtained using
vides more detailed information on the distribution of the DDHT model within days or even hours (mostly con-
electronic states with respect to different energy levels, sumed by matrix diagonalization). As expected, the pre-
chemical elements, atoms, and orbitals, supplementing dicted band structures of the two small-angle bP twisted
the picture of electronic structures other than dispersion supercells, as shown in Fig. 5b, c, clearly display the cor-
relations demonstrated above. related flat-band feature in their top valence states, of-
At the current stage, the above demonstrations of our fering a new opportunity to study flat-band physics and
database in accurately predicting DFT Hamiltonians and properties in an ab initio manner. Obtaining the band
electronic structures cover all 124 + 5 types of twisted bi- structures of the two bP twisted structures with thou-
layer materials considered, which indicates its high gen- sands of atoms using ab initio DFT methods consumes
eralizability and reliability as a DFT-level deep-learning unacceptable computational resources, however, they can
model database that offers a wide set of DeepH-E3 mod- be obtained using the DDHT model within days or even
els targeting the prediction Hamiltonians of twisted ma- hours (mostly spent on matrix diagonalization).
terials, effectively bypassing the expensive first-principles For the twisted bilayer bP structure with twist angle
calculations. The research into twisted materials and re- θ = 2.36◦ , our DDHT model for bP predicts the presence
lated areas can be greatly assisted and accelerated by our of two stand-alone flat bands (Fig. 5b) separated by 8.4
database, as well as the predicted results, through offer- meV at the valence band top. The band width of the
ing guidance and reference prior to costly investigations first (marked by red) and the second valence flat band
and screening for more exotic properties. (marked by blue) are 7.7 and 1.0 meV. Their small band
width enables exotic correlated states when they are par-
tially occupied upon hole doping. Moreover, the closely
Prediction of Flat Bands located flat bands with respect to energy make the corre-
lated states from different flat bands accessible through
While our current demonstrations of predicting DFT modulation of doping density in a small range. The real-
Hamiltonians and electronic structures for twisted struc- space projections of these two flat bands (5e) are quite
tures are limited to θ values as low as 13.17◦ for the localized and well separated by half the moiré periodic-
majority of the considered materials, the remarkable gen- ity, which also support the strong correlated states. The
eralizability of our database extends the capability to ex- flat bands and their real-space projections of twisted bP
plore twisted structures with arbitrarily small θ angles—a of smaller θ = 1.59◦ (Fig. 5c, f) are also predicted by
 7

a bP, θ≈10.9°, 228 atoms c TaS2, θ≈21.8°, 42 atoms e Bi2Se3-Bi2Te3, θ≈21.8°, 70 atoms

b bP, θ≈8.9°, 364 atoms d TaS2, θ≈13.2°, 114 atoms f Bi2Se3-Bi2Te3, θ≈13.2°, 190 atoms

FIG. 4. The DFT-calculated band structures (red curves) and predicted band structures (blue dots) for twisted (a, b) bP with
θ = 10.87◦ and 8.49◦ , and (c, d) 2H-phase TaS2 , and (e, f ) heterobilayer Bi2 Se3 /Bi2 Te3 with θ = 21.79◦ and 13.17◦ . The twist
angles of bP are different due to the consideration of lattice compression in constructing twisted supercells from rectangular
lattices. The accuracy of both the DDHT models and predicted electronic structures is demonstrated by almost identical bands
from predictions and DFT calculations.

our DDHT model. In Fig. 5c, multiple stand-alone flat are not limited to predicting only such structures. In
valence bands are identified in energy range between ± fact, our previous examples on the band-structure predic-
0.2 eV and have band width around 1.0 meV, potentially tion of relaxed bP twisted bilayers demonstrated that the
giving rise to more pronounced correlated phases than in prediction accuracy is not significantly affected by struc-
larger-angle cases. tural relaxation. This is mainly because the atomic dis-
The accurate and efficient predictions of flat bands and placements induced by relaxation are generally below or
their real-space projections pave the way for discover- around 0.2 Å, the resulting interatomic distances between
ing new flat-band materials and future experimental and neighboring atoms after relaxation are still within the
theoretical investigations into the correlated physics in coverage of our training datasets (eg, Fig. S2). By using
twisted bP and other twisted materials. the same model that predicts for ion-clamped TBG, the
DeepH-E3-predicted energy bands of relaxed magic-angle
TBG fit closely to its DFT-calculated ones [22], which
DISCUSSION further support this argument. The obtained MAEs in
both cases are in the sub-meV range, indicating that the
While all the previous discussions have mainly focused models in the DDHT can capture the structural relax-
on ion-clamped atomic structures of twisted materials, it ation accurately. Therefore, we can conclude that the
should be noted that the DeepH-E3 models in the DDHT DDHT models are applicable to a wide range of twisted
 8

a 12,500 c
600
atoms/unit cell
10,000 Infer time 500
atoms/unit cell

Infer time (s)

Energy (eV)
7,500 400

300
5,000
200

2,500
100

0 0
7.04

5.97
5.18

3.89
3.54
3.24
2.99
2.78
2.51
2.36
2.10
1.90
1.77
1.59
Twist angle (°)
b 250 d
HOMO bandwidth
200
Average bandwidth
Bandwidth (meV)

Energy (eV)
150

100

50

0
7.04

5.97
5.18

3.89
3.54
3.24
2.99
2.78
2.51
2.36
2.10
1.90
1.77
1.59

Twist angle (°)

FIG. 5. Prediction of flat bands in electronic structures of twisted bP systems. (a) shows the number of atoms per unit
cell of twisted structures (blue bars) and the associated time of Hamiltonian inference using the trained model in DDHT for
different twist angles. The inference time increases almost linearly with the number of atoms. (b) plots the bandwidth of
highest occupied band, average bandwidth and standard error of 10 bands, which contains 5 highest valance bands and 5 lowest
conduction bands, against the twist angle. The angle values on x-axis of (a, b) use logarithmic scale for clear demonstration.
Predicted band structures and real-space flat-band projections of twisted bP structures with (c) θ = 2.36◦ and (d) 1.59◦ as
examples. The top two flat valence bands are outlined by red and blue and their projections in real space are presented by
transparent regions in the same color. The twisted atomic structures are schematically displayed in the background.

materials structures, whether they are ion-clamped or re- ously demonstrated by the predictions of twisted hetero-
laxed. bilayer structures including graphene/BN, MoS2 /MoSe2 ,
WS2 /WSe2 , Bi2 Se3 /Bi2 Te3 , and MoTe2 /WSe2 , and
Straightforwardly, the models in the DDHT are highly can potentially sustain in predicting twisted multi-
versatile in predicting electronic structures of twisted heterolayers.
multilayer structures, which can be treated as combina-
tions of twisted bilayers consist of the same or different The DDHT currently covers only a fraction of stable
materials. For instance, in a twisted trilayer structure, 2D materials (124 out of about 2132 non-magnetic mate-
where layers 1, 2, and 3 are arranged in a top-to-bottom rials in the C2DB accessed at June 19, 2022), it will cover
order, layer 1-2 and layer 2-3 constitute two twisted bi- more non-magnetic materials in the near-future develop-
layers. As a result, the Hamiltonian and electronic struc- ment, by using the optimal scheme of dataset prepara-
ture of this twisted trilayer can be accurately predicted tion and DeepH-E3 model training discussed in this ar-
using the same models as in the bilayer cases. This argu- ticle. Especially, we will take priority at offering models
ment can be extended to twisted hetero-multilayer struc- involving materials which users of our DDHT concern
tures where the constituent layers are all different ma- most. Although the models in the DDHT can predict ar-
terials. In this case, multiple models trained by heter- bitrary twist angles, the diagonalization to obtain band
obilayer training datasets are required for accurate pre- structure, DOS or other properties will consume certain
dictions. The accuracy of these models has been previ- amount of computational resource. Thus we allow users
 9

to upload their results help to enrich the database. In models and related files like predicted DFT Hamiltonian
conclusion, the DDHT will be updated with predictions matrices, eigenvalues and eigenvectors, band structures,
for more twist angles, especially for twisted materials and density of states, and identified flat bands. Further, by
twist angles that are currently of general research inter- screening the database, we list ultra flat bands in small-
est. Additionally, the magnetic materials (about 464 sta- angle twisted bP and other 19 types of large-angle twisted
ble in the C2DB) present a unique challenge due to their materials. The accuracy of our DDHT models and pre-
spin degree of freedom, which requires the incorporation dictions is demonstrated with comparable MAEs and en-
of both spin moments and crystal structures as input ergy bands to respective DFT results. Our DDHT cov-
in learning the neural network representations of their ers twisted structures of a significant number of 2D ma-
Hamiltonians. This can be achieved through the uti- terials and holds solid universality over a wide range of
lization of the recently developed xDeepH method [33], twist angles down to about 1◦ , showing the reliability and
which can predict DFT Hamiltonians of magnetic twisted universality of DeepH method and corresponding high-
materials. Therefore, in the future development of the throughput NN parameter strategy, offering DFT-level
DDHT, xDeepH models and corresponding predictions knowledge prior to deep investigations and serving as a
of magnetic materials can be included. In our vision, the handy tool for researchers in the field of twistronics. An-
DDHT has the potential to cover all stable 2D materials other crucial contribution of this work if to perform the
in the future and serve as a handy tool for researchers in new paradigm of building a material database, which is a
the emerging field of twistronics. collection of useful models applied to series of materials
The predicted results in the DDHT are not limited instead of single material and can be seen as a extensible
to Hamiltonians and electronic structures, the identi- MoE model. In the future development, our DDHT can
fications of topology in twisted materials will also be be expanded to cover all stable 2D materials and incor-
included in future development. This will be achieved porate magnetism and topology, satisfying the growing
by determining the irreducible representations of va- general interest toward exotic properties in twisted ma-
lence eigenstates at high-symmetry points in momentum terials.
space under various symmetry operations and comparing
them with the elementary band representations (EBRs)
in the topological quantum chemistry [34–36]. If these METHODS
irreducible representations of a material are not linear
combinations of EBRs, it would be identified as having Preparation of Training Datasets
non-trivial topology. Based on the predicted eigenstates
by the DDHT models, theoretically, the band topology The bilayer unit cell for each type of 2D material in
of twisted materials of any chemical constitutions and our database is fully relaxed using Viena ab initio Simu-
twist angles can be accurately and efficiently determined; lation Package (VASP) [28, 38] to give the most optimal
meanwhile, the berry curvature may also be calculated lattice constants, stacking orders, and interlayer spac-
to explore exotic properties regarding topology. Note ings. The interactions between ions and electrons and
that the classification of topology by topological quan- the exchange correlations are treated by the projected-
tum chemistry holds effective in the case of the generally- augmented wave method [39, 40] and the Perdew-Burke-
existed flat bands in twisted materials [37]. In this way, Ernzerhof functional [41]. The interaction is corrected
our DDHT may promisingly find all possible topological by DFT-D3 method [42, 43]. The convergence criteria
twisted materials and provide opportunities in explor- for energy and force are 1 × 10−7 eV and 1 × 10−3 eV/Å.
ing the currently scarce twist angle-dependent topologi- The vacuum space between two periodic images along the
cal transitions. direction normal to the 2D plane is set to be more than 12
Å. The k-point mesh is 11×11×1 using the Monkhorst-
Pack scheme [44].
SUMMARY The atomic structures in the training dataset for each
2D material in the DDHT are prepared as a series of 3×3
In summary, by using the deep-learning method tar- non-twisted bilayer supercells (made from the optimized
geting at DFT Hamiltonians implemented in the DeepH- primitive bilayer unit cells) with typically 24×24 uni-
E3 code, we train the deep neural network models for formly shifted interlayer stackings along two in-plane lat-
124(5) types of homo(hetero)bilayer 2D materials and tice vectors followed by random perturbations to atomic
use them to predict DFT Hamiltonians, as well as elec- positions (up to 0.1 Å) along three respective Cartesian
tronic structures, of their twisted structures at DFT ac- axes. In total, the training dataset for each material has
curacy. Based on these well-trained models capable of typically 576 perturbed and shifted 3 × 3 supercell struc-
doing accurate predictions, we construct a deep-learning tures.
model database for Hamiltonian prediction of twisted The Hamiltonian for each shifted and perturbed
materials named DDHT, which consists of DeepH-E3 NN atomic structure in the training datasets is calculated by
 10

first-principles density functional theory (DFT) simula- the locality of local-atomic basis and the eigenvalues and
tion code OpenMX [45, 46] using pseudo atomic orbital eigenvectors are calculated by efficient matrix diagonal-
basis, norm-conserving pseudo potential [47], and PBE ization using the Pardiso package [49, 50] of Julia code.
exchange functional [41]. The energy convergence crite-
ria, cut-off energy, and Γ-centered k-point mesh are set
as 4 × 10−8 Hartree, 300 Ry, and no less than 3 × 3 × 1 in DATA AVAILABILITY
the self-consistent calculations, respectively. The DFT
Hamiltonians and band structures used for evaluating
The DeepH-E3 models of all 124 homobilayer and 5
the DDHT accuracy are calculated by OpenMX code us-
heterobilayer materials and their twisted structures of
ing similar settings. The training datasets, including the
large θs will be available at the DDHT website.
DFT Hamiltonians and the corresponding structures, are
processed prior to the training. The DFT Hamiltonians
are decomposed into small matrix blocks Hij that cor-
responds to local atomic structures Rij within the cut- ACKNOWLEDGMENTS
off radius Rc , and these Rij are represented by crystal
graphs where the on-site atoms and the distances to their This work was supported by the Basic Science Cen-
neighbors within Rc are treated by vertex and edge fea- ter Project of NSFC (grant no. 2388201), the Na-
tures. tional Natural Science Foundation of China (grant no.
12334003), the National Science Fund for Distinguished
Young Scholars (grant no. 12025405) and the Min-
Training and Prediction istry of Science and Technology of China (grant no.
2023YFA1406400). R.X. acknowledges the founding sup-
The deep neural network models in the DDHT are port from the Fellowship of China Postdoctoral Science
trained by the DeepH-E3 code [21, 22] with Euclidean Foundation (Grant No. 2021TQ0187). to be updated
symmetry of local coordinate transformation satisfied
by an E(3)-equivariant framework [48]. Details of the
DeepH-E3 code used in the DDHT are summarized as AUTHOR CONTRIBUTIONS
follows. The representation for the initial and the in-
termediate vertex (edge) features is 64×0e and 64×0e + T.B. and R.X. contributed equally to this project.
32×1o + 16×2e + 8×3o + 8×4e, respectively, where the
first and the second number (separated by ”×”) in each
term (connected by ”+”) stands respectively for the num-
COMPETING INTERESTS
ber of vectors and the carried representation of spherical
harmonics l, while the last letter e (o) denotes even (odd)
parity. Spherical harmonics with l = 0 to 4 and l = 0 to The authors declare no competing of interests.
5 are used for materials without and with spin-orbit cou-
pling respectively. During the model training, we employ
an optimized learning scheme where the training, valida-
tion, and testing set take respectively 80%, 15%, and 5% ∗
[email protected]
of the total dataset, the batch size takes the value of †
[email protected]
2, and the learning rate starts with 0.002 and reduces by ‡
[email protected]
half when validation loss plateau (in this case, decreasing [1] Y. Cao, V. Fatemi, A. Demir, S. Fang, S. L. Tomarken,
less than 5% after every 120 epochs). The training takes J. Y. Luo, J. D. Sanchez-Yamagishi, K. Watanabe,
at least 1,000 epochs and is terminated upon the vali- T. Taniguchi, E. Kaxiras, R. C. Ashoori, and P. Jarillo-
dation loss drops below 1 × 10−7 for its averaged value Herrero, Correlated insulator behaviour at half-filling
in magic-angle graphene superlattices, Nature 556, 80
or 1 × 10−5 for its maximum value, ensuring the high (2018).
accuracy and generalizability of DeepH-E3 models. Sim- [2] Y. Cao, V. Fatemi, S. Fang, K. Watanabe, T. Taniguchi,
ilar loss values are given in evaluation of these models in E. Kaxiras, and P. Jarillo-Herrero, Unconventional super-
testing sets upon the stopping of model training. conductivity in magic-angle graphene superlattices, Na-
The Hamiltonian matrices are predicted using only ture 556, 43 (2018).
atomic structures of twisted bilayer materials, corre- [3] X. Sun, M. Suriyage, A. R. Khan, M. Gao, J. Zhao,
B. Liu, M. M. Hasan, S. Rahman, R.-s. Chen, P. K. Lam,
sponding overlap matrices, and trained DeepH-E3 mod-
et al., Twisted van der waals quantum materials: Funda-
els as input. Based on the atomic structures and local- mentals, tunability, and applications, Chemical Reviews
atomic basis, the overlap matrices are directly obtained (2024).
by a modified OpenMX code without conducting DFT [4] T. Song, Q.-C. Sun, E. Anderson, C. Wang, J. Qian,
calculations. The HDeepH-E3 matrices are sparse due to T. Taniguchi, K. Watanabe, M. A. McGuire, R. Stöhr,
 11

D. Xiao, T. Cao, J. Wrachtrup, and X. Xu, Direct visu- [18] A. W. Senior, R. Evans, J. Jumper, J. Kirkpatrick,
alization of magnetic domains and moiré magnetism in L. Sifre, T. Green, C. Qin, A. Žı́dek, A. W. R. Nel-
twisted 2d magnets, Science 374, 1140 (2021). son, A. Bridgland, H. Penedones, S. Petersen, K. Si-
[5] Y. Xu, A. Ray, Y.-T. Shao, S. Jiang, K. Lee, D. We- monyan, S. Crossan, P. Kohli, D. T. Jones, D. Sil-
ber, J. E. Goldberger, K. Watanabe, T. Taniguchi, D. A. ver, K. Kavukcuoglu, and D. Hassabis, Improved protein
Muller, K. F. Mak, and J. Shan, Coexisting ferromag- structure prediction using potentials from deep learning,
netic–antiferromagnetic state in twisted bilayer CrI3, Na- Nature 577, 706 (2020).
ture Nanotechnology 17, 143 (2021). [19] L. Zhang, J. Han, H. Wang, R. Car, and W. E, Deep
[6] M. V. Stern, Y. Waschitz, W. Cao, I. Nevo, K. Watanabe, potential molecular dynamics: A scalable model with the
T. Taniguchi, E. Sela, M. Urbakh, O. Hod, and M. B. accuracy of quantum mechanics, Physical Review Letters
Shalom, Interfacial ferroelectricity by van der waals slid- 120, 143001 (2018).
ing, Science 372, 1462 (2021). [20] J. Behler and M. Parrinello, Generalized neural-network
[7] K. Yasuda, X. Wang, K. Watanabe, T. Taniguchi, and representation of high-dimensional potential-energy sur-
P. Jarillo-Herrero, Stacking-engineered ferroelectricity in faces, Physical review letters 98, 146401 (2007).
bilayer boron nitride, Science 372, 1458 (2021). [21] H. Li, Z. Wang, N. Zou, M. Ye, R. Xu, X. Gong,
[8] T. Devakul, V. Crépel, Y. Zhang, and L. Fu, Magic in W. Duan, and Y. Xu, Deep-learning density functional
twisted transition metal dichalcogenide bilayers, Nature theory Hamiltonian for efficient ab initio electronic-
communications 12, 6730 (2021). structure calculation, Nature Computational Science 2,
[9] C. Jin, E. C. Regan, A. Yan, M. I. B. Utama, D. Wang, 367 (2022).
S. Zhao, Y. Qin, S. Yang, Z. Zheng, S. Shi, K. Watanabe, [22] X. Gong, H. Li, N. Zou, R. Xu, W. Duan, and Y. Xu,
T. Taniguchi, S. Tongay, A. Zettl, and F. Wang, Obser- General framework for E(3)-equivariant neural network
vation of moiré excitons in WSe2/WS2 heterostructure representation of density functional theory Hamiltonian,
superlattices, Nature 567, 76 (2019). Nature Communications 14, 2848 (2023).
[10] K. Tran, G. Moody, F. Wu, X. Lu, J. Choi, K. Kim, [23] A. Jain, S. P. Ong, G. Hautier, W. Chen, W. D. Richards,
A. Rai, D. A. Sanchez, J. Quan, A. Singh, J. Emb- S. Dacek, S. Cholia, D. Gunter, D. Skinner, G. Ceder,
ley, A. Zepeda, M. Campbell, T. Autry, T. Taniguchi, and K. A. Persson, Commentary: The materials project:
K. Watanabe, N. Lu, S. K. Banerjee, K. L. Silverman, A materials genome approach to accelerating materials
S. Kim, E. Tutuc, L. Yang, A. H. MacDonald, and X. Li, innovation, APL Materials 1, 011002 (2013).
Evidence for moiré excitons in van der waals heterostruc- [24] M. Hellenbrandt, The inorganic crystal structure
tures, Nature 567, 71 (2019). database (icsd)—present and future, Crystallography Re-
[11] E. M. Alexeev, D. A. Ruiz-Tijerina, M. Danovich, M. J. views 10, 17 (2004).
Hamer, D. J. Terry, P. K. Nayak, S. Ahn, S. Pak, J. Lee, [25] S. Haastrup, M. Strange, M. Pandey, T. Deilmann, P. S.
J. I. Sohn, M. R. Molas, M. Koperski, K. Watanabe, Schmidt, N. F. Hinsche, M. N. Gjerding, D. Torelli, P. M.
T. Taniguchi, K. S. Novoselov, R. V. Gorbachev, H. S. Larsen, A. C. Riis-Jensen, J. Gath, K. W. Jacobsen, J. J.
Shin, V. I. Fal’ko, and A. I. Tartakovskii, Resonantly hy- Mortensen, T. Olsen, and K. S. Thygesen, The Compu-
bridized excitons in moiré superlattices in van der waals tational 2D Materials Database: High-throughput mod-
heterostructures, Nature 567, 81 (2019). eling and discovery of atomically thin crystals, 2D Mate-
[12] Y. Shimazaki, I. Schwartz, K. Watanabe, T. Taniguchi, rials 5, 042002 (2018).
M. Kroner, and A. Imamoğlu, Strongly correlated elec- [26] M. N. Gjerding, A. Taghizadeh, A. Rasmussen, S. Ali,
trons and hybrid excitons in a moiré heterostructure, Na- F. Bertoldo, T. Deilmann, N. R. Knøsgaard, M. Kruse,
ture 580, 472 (2020). A. H. Larsen, S. Manti, T. G. Pedersen, U. Petralanda,
[13] E. C. Regan, D. Wang, C. Jin, M. I. B. Utama, B. Gao, T. Skovhus, M. K. Svendsen, J. J. Mortensen, T. Olsen,
X. Wei, S. Zhao, W. Zhao, Z. Zhang, K. Yumigeta, and K. S. Thygesen, Recent progress of the Computa-
M. Blei, J. D. Carlström, K. Watanabe, T. Taniguchi, tional 2D Materials Database (C2DB), 2D Materials 8,
S. Tongay, M. Crommie, A. Zettl, and F. Wang, Mott and 044002 (2021).
generalized wigner crystal states in WSe2/WS2 moiré su- [27] R. A. Jacobs, M. I. Jordan, S. J. Nowlan, and G. E.
perlattices, Nature 579, 359 (2020). Hinton, Adaptive mixtures of local experts, Neural com-
[14] L. Xian, D. M. Kennes, N. Tancogne-Dejean, putation 3, 79 (1991).
M. Altarelli, and A. Rubio, Multiflat bands and strong [28] G. Kresse and J. Furthmüller, Efficiency of ab-initio total
correlations in twisted bilayer boron nitride: Doping- energy calculations for metals and semiconductors using
induced correlated insulator and superconductor, Nano a plane-wave basis set, Computational Materials Science
Letters 19, 4934 (2019). 6, 15 (1996).
[15] D. M. Kennes, L. Xian, M. Claassen, and A. Rubio, One- [29] G. Kresse and J. Furthmüller, Efficient iterative schemes
dimensional flat bands in twisted bilayer germanium se- for ab initio total-energy calculations using a plane-wave
lenide, Nature Communications 11, 1124 (2020). basis set, Physical review B 54, 11169 (1996).
[16] D. M. Kennes, M. Claassen, L. Xian, A. Georges, A. J. [30] T. Xie and J. C. Grossman, Crystal graph convolutional
Millis, J. Hone, C. R. Dean, D. N. Basov, A. N. Pa- neural networks for an accurate and interpretable predic-
supathy, and A. Rubio, Moiré heterostructures as a tion of material properties, Physical Review Letters 120,
condensed-matter quantum simulator, Nature Physics 145301 (2018).
17, 155 (2021). [31] S. Manzeli, D. Ovchinnikov, D. Pasquier, O. V. Yazyev,
[17] S. Carr, S. Fang, and E. Kaxiras, Electronic-structure and A. Kis, 2d transition metal dichalcogenides, Nature
methods for twisted moiré layers, Nature Reviews Mate- Reviews Materials 2, 17033 (2017).
rials 5, 748 (2020). [32] V. L. Deringer, M. A. Caro, and G. Csányi, A general-
purpose machine-learning force field for bulk and nanos-
 12

tructured phosphorus, Nature Communications 11, 5461

(2020).
[33] H. Li, Z. Tang, X. Gong, N. Zou, W. Duan, and Y. Xu,
Deep-learning electronic-structure calculation of mag-
netic superstructures, Nature Computational Science 3,
321 (2023).
[34] B. Bradlyn, L. Elcoro, J. Cano, M. G. Vergniory,
Z. Wang, C. Felser, M. I. Aroyo, and B. A. Bernevig,
Topological quantum chemistry, Nature 547, 298 (2017).
[35] M. G. Vergniory, L. Elcoro, C. Felser, N. Regnault, B. A.
Bernevig, and Z. Wang, A complete catalogue of high-
quality topological materials, Nature 566, 480 (2019).
[36] M. G. Vergniory, B. J. Wieder, L. Elcoro, S. S. P. Parkin,
C. Felser, B. A. Bernevig, and N. Regnault, All topolog-
ical bands of all nonmagnetic stoichiometric materials,
Science 376, eabg9094 (2022).
[37] D. Călugăru, A. Chew, L. Elcoro, Y. Xu, N. Regnault, Z.-
D. Song, and B. A. Bernevig, General construction and
topological classification of crystalline flat bands, Nature
Physics 18, 185 (2021).
[38] G. Kresse and J. Furthmüller, Efficient iterative schemes
for ab initio total-energy calculations using a plane-wave
basis set, Physical Review B 54, 11169 (1996).
[39] P. E. Blöchl, Projector augmented-wave method, Physi-
cal Review B 50, 17953 (1994).
[40] G. Kresse and D. Joubert, From ultrasoft pseudopoten-
tials to the projector augmented-wave method, Physical
Review B 59, 1758 (1999).
[41] J. P. Perdew, K. Burke, and M. Ernzerhof, Generalized
gradient approximation made simple, Physical Review
Letters 77, 3865 (1996).
[42] S. Grimme, S. Ehrlich, and L. Goerigk, Effect of the
damping function in dispersion corrected density func-
tional theory, Journal of Computational Chemistry 32,
1456 (2011).
[43] S. Grimme, J. Antony, S. Ehrlich, and H. Krieg, A con-
sistent and accurate ab initio parametrization of den-
sity functional dispersion correction (dft-d) for the 94
elements h-pu, The Journal of Chemical Physics 132,
154104 (2010).
[44] H. J. Monkhorst and J. D. Pack, Special points for
brillouin-zone integrations, Physical Review B 13, 5188
(1976).
[45] T. Ozaki, Variationally optimized atomic orbitals for
large-scale electronic structures, Phys. Rev. B 67, 155108
(2003).
[46] T. Ozaki and H. Kino, Numerical atomic basis orbitals
from h to kr, Phys. Rev. B 69, 195113 (2004).
[47] I. Morrison, D. M. Bylander, and L. Kleinman, Nonlo-
cal hermitian norm-conserving vanderbilt pseudopoten-
tial, Phys. Rev. B 47, 6728 (1993).
[48] M. Geiger and T. Smidt, e3nn: Euclidean neu-
ral networks, arXiv preprint arXiv:2207.09453
10.48550/arXiv.2207.09453 (2022).
[49] C. Alappat, A. Basermann, A. R. Bishop, H. Fehske,
G. Hager, O. Schenk, J. Thies, and G. Wellein, A recur-
sive algebraic coloring technique for hardware-efficient
symmetric sparse matrix-vector multiplication, ACM
Trans. Parallel Comput. 7, 37 (2020).
[50] M. Bollhöfer, A. Eftekhari, S. Scheidegger, and
O. Schenk, Large-scale sparse inverse covariance matrix
estimation, SIAM Journal on Scientific Computing 41,
A380 (2019).