R ES E A RC H

◥ techniques to attack these problems, artificial

RESEARCH ARTICLE neural networks play a prominent role (16). They
can perform exceedingly well in a variety of con-
texts ranging from image and speech recognition
MANY-BODY PHYSICS (17) to game playing (18). Very recently, appli-
cations of neural networks to the study of phy-

Solving the quantum many-body sical phenomena have been introduced (19–23).
These have so far focused on the classification
of complex phases of matter, when exact sampling
problem with artificial of configurations from these phases is possible.
The challenging goal of solving a many-body

neural networks problem without prior knowledge of exact sam-

ples is nonetheless still unexplored, and the po-
tential benefits of artificial intelligences in this
Giuseppe Carleo1* and Matthias Troyer1,2 task are, at present, substantially unknown.
Therefore, it is of fundamental and practical in-
The challenge posed by the many-body problem in quantum physics originates from the terest to understand whether an artificial neural
difficulty of describing the nontrivial correlations encoded in the exponential complexity network can modify and adapt itself to describe
of the many-body wave function. Here we demonstrate that systematic machine learning of and analyze such a quantum system. This abil-

Downloaded from http://science.sciencemag.org/ on February 9, 2017

the wave function can reduce this complexity to a tractable computational form for some ity could then be used to solve the quantum
notable cases of physical interest. We introduce a variational representation of quantum many-body problem in regimes that have tra-
states based on artificial neural networks with a variable number of hidden neurons. ditionally been inaccessible to existing exact nu-
A reinforcement-learning scheme we demonstrate is capable of both finding the ground merical approaches.
state and describing the unitary time evolution of complex interacting quantum systems. Here we introduce a representation of the wave
Our approach achieves high accuracy in describing prototypical interacting spins models in function in terms of artificial neural networks
one and two dimensions. specified by a set of internal parameters W. We

T
present a stochastic framework for reinforce-
he wave function Y is a fundamental ob- a large number of unexplored regimes exist, in- ment learning of the parameters W, allowing for
ject in quantum physics and possibly the cluding many open problems. These encompass the best possible representation of both ground-
hardest to grasp in the classical world. Y fundamental questions ranging from the dyna- state and time-dependent physical states of a
is a monolithic mathematical quantity that mical properties of high-dimensional systems given quantum Hamiltonian H. The parame-
contains all of the information on a quan- (11, 12) to the exact ground-state properties of ters of the neural network are then optimized
tum state, be it a single particle or a complex strongly interacting fermions (13, 14). At the heart (trained, in the language of neural networks),
molecule. In principle, an exponential amount of this lack of understanding lies the difficulty either by static variational Monte Carlo (VMC)
of information is needed to fully encode a ge- in finding a general strategy to reduce the ex- sampling (24) or time-dependent VMC (25, 26),
neric many-body quantum state. However, wave ponential complexity of the full many-body wave when dynamical properties are of interest. We
functions representing many physical many-body function down to its most essential features (15). validate the accuracy of this approach by study-
systems can be characterized by an amount of In a much broader context, the problem re- ing the Ising and Heisenberg models in both one
information much smaller than the maximum sides in the realm of dimensional reduction and and two dimensions. The power of the neural-
capacity of the corresponding Hilbert space. A feature extraction. Among the most successful network quantum states (NQS) is demonstrated,
limited amount of quantum entanglement and
a small number of physical states in such sys-
tems enable modern approaches to solve the
many-body Schrödinger’s equation with a limited
amount of classical resources.
h1 h2 h3 Hidden Layer hM
Numerical approaches directly relying on the
wave function can either sample a finite num-
ber of physically relevant configurations or per-
form an efficient compression of the quantum
state. Stochastic approaches, like quantum Monte
Carlo (QMC) methods, belong to the first cat-
egory and rely on probabilistic frameworks typ-
ically demanding a positive semidefinite wave
function (1–3). Compression approaches instead
rely on efficient representations of the wave func-
tion, such as in terms of matrix product states
(MPS) (4–6) or more general tensor networks
(7–9). However, examples of systems in which

σ1z σ2z σ3z z

σN
existing approaches fail are numerous, mostly
owing to the sign problem in QMC (10) and Visible Layer
to the inefficiency of current compression ap-
proaches in high-dimensional systems. As a result,
despite the notable success of these methods,
1
Fig. 1. Artificial neural network encoding a many-body quantum state of N spins. A restricted
Theoretical Physics, ETH Zurich, 8093 Zurich, Switzerland.
2 Boltzmann machine architecture that features a set of N visible artificial neurons (yellow dots) and a
Quantum Architectures and Computation Group, Microsoft
Research, Redmond, WA 98052, USA. set of M hidden neurons (gray dots) is shown. For each value of the many-body spin configuration
*Corresponding author. Email: [email protected] S ¼ ðs1z ; s2z ; … ; sNz Þ, the artificial neural network computes the value of the wave function YðSÞ.

Carleo et al., Science 355, 602–606 (2017) 10 February 2017 1 of 4

R ES E A RC H | R E S EA R C H A R T I C LE

obtaining state-of-the-art accuracy in both ground- convenient point of the NQS representation is many-body state is unknown, and it is typically
state and out-of-equilibrium dynamics. that it can be formulated in a way that conserves found upon solving either the static Schrödinger
some specific symmetries. For example, lattice trans- equation HjYi ¼ EjYi or the time-dependent one
Neural-network quantum states lation symmetry can be used to reduce the number HjYðtÞi ¼ i dtd jYðtÞi for a given Hamiltonian H.
Consider a quantum system with N discrete- of variational parameters of the NQS ansatz, in the In the absence of samples drawn according to the
valued degrees of freedom S ¼ ðS 1 ; S 2 ; … ; S N Þ, spirit of shift-invariant RBMs (30, 31). Concretely, exact wave function, supervised learning of Y is
which may be spins, bosonic occupation numbers, for integer hidden-variable density a = 1,2,…, the therefore not a viable option. Instead, we derive
or similar. The many-body wave function is a map- weight matrix takes the form of feature filters a consistent reinforcement learning approach in
ðf Þ
ping of the N-dimensional set S to (exponentially Wj for f ∈½1; a. These filters have a total of aN which either the ground-state wave function or
many) complex numbers that fully specify the variational elements in lieu of the aN 2 elements of the time-dependent one is learned on the basis
amplitude and the phase of the quantum state. the asymmetric case (see supplementary materials). of feedback from variational principles.
The point of view we take here is to interpret Given a general expression for the quantum
the wave function as a computational black box many-body state, we are now left with the task of Ground state
which, given an input many-body configuration solving the many-body problem by using machine To demonstrate the accuracy of the NQS in
S, returns a phase and an amplitude according learning to optimize the network parameters W. the description of complex many-body quantum
to YðSÞ. Our goal is to approximate this compu- In the most interesting applications, the exact states, we first focus on the goal of finding the
tational black box with a neural network, trained
to best represent YðSÞ. Different possible choices
for the artificial neural-network architectures Ising 1D Heisenberg 1D

Downloaded from http://science.sciencemag.org/ on February 9, 2017

have been proposed to solve specific tasks, and
the best architecture to describe a many-body
quantum system may vary from one case to an- W
other. For the sake of concreteness, we henceforth
specialize our discussion to restricted Boltzmann
machine (RBM) architectures and apply them to
W
describe spin-½ quantum systems. In this case,
RBM artificial networks are constituted by one
visible layer of N nodes, corresponding to the
physical spin variables in a chosen basis (e.g., W
S ¼ sz1 ; …; szN ) and a single hidden layer of M
auxiliary spin variables ðh1 ; …; hM Þ (Fig. 1). This
description corresponds to a variational expres-
sion for the quantum states W
X X X
X aj szj þ bi hi þ Wij hi szj
YM ðS; WÞ ¼ e j i
ij

fhi g

where hi ¼ f−1; 1g is a set of M hidden spin varia-

0.15 0.00 0.15
bles and the network parameters W ¼ fa; b; W g
fully specify the response of the network to a Heisenberg 2D
given input state S. Because this architecture fea-
tures no intralayer interactions, the hidden varia-
bles can be explicitly traced out, and the wave
P z
aj s
j
W (1) W (2) W (3) W (4)
function reads YðS; WÞ ¼ e  PM i¼1 Fi ðSÞ,
j

P
where Fi ðSÞ ¼ 2cosh½bi þ W ij sj . The network
z
j
weights are, in general, to be taken complex- W (5) W (6) W (7) W (8)
valued to provide a complete description of both
the amplitude and the phase of the wave function.
The mathematical foundations for the ability
W (9) W (10) W (11) W (12)
of NQS to describe intricate many-body wave
functions are the established representability
theorems (27–29), which guarantee the existence
of network approximates of sufficiently smooth W (13) W (14) W (15) W (16)
and regular high-dimensional functions. If these
conditions are satisfied by the many-body wave
function, we can reasonably expect the NQS form 0.3 0.0 0.3
to be a sensible choice. One of the practical ad-
vantages of this representation is that its quality Fig. 2. Neural-network representation of the many-body ground states. Results for prototypical
can, in principle, be systematically improved by spin models in one and two dimensions are shown. In the top group of panels, we show the feature maps for
increasing the number of hidden variables. The the 1D transverse-field Ising (TFI) model at the critical point h = 1, as well as for the antiferromagnetic
number M (or, equivalently, the density a = M/N) Heisenberg (AFH) model. In both cases, the hidden-unit density is a = 4 and the lattices comprise 80 sites.
ðf Þ
then plays a role analogous to the bond dimension Each horizontal colormap shows the values that the fth feature map W j takes on the jth lattice site (horizontal
for the MPS. However, the correlations induced axis, broadened along the vertical direction for clarity). In the bottom group of panels, we show the feature
by the hidden units are intrinsically nonlocal in maps for the 2D Heisenberg model on a square lattice, for a = 16. In this case, the horizontal (or vertical) axis
space and are therefore well suited to describe of the colormaps corresponds to the x (or y) coordinates on a 10-by-10 square lattice. Each of the feature
quantum systems in arbitrary dimension. Another maps acts as an effective filter on the spin configurations, capturing the most important quantum correlations.

Carleo et al., Science 355, 602–606 (2017) 10 February 2017 2 of 4

R ES E A RC H | R E S EA R C H A R T I C LE

best neural-network representation of the un- 10−2 10−2

known ground state of a given Hamiltonian H. A B 10−2 C
In this context, reinforcement learning is realized h 10−3
through minimization of the expectation value 10−4
of the energy EðWÞ ¼ hYM jHjYM i=hYM jYM i 10−4
10−3


with respect to the network weights W. In the −6 h
10 −5
stochastic setting, this is achieved with an iterative 10
scheme. At each iteration k, a Monte Carlo sampl- h
ing of jYM ðS; W k Þj2 is realized for a given set of 10−8 10−6 10−4
1 2 4 1 2 4 1 2 4 8 16 32
parameters W k . At the same time, stochastic esti- α α α
mates of the energy gradient are obtained. These
are then used to propose a next set of weights Fig. 3. Finding the many-body ground-state energy with neural-network quantum states (NQS).
W kþ1 with an improved gradient-descent optimi- The error of the NQS ground-state energy relative to the exact value is shown for several test cases.
zation (32). The overall computational cost of Arbitrary precision on the ground-state energy can be obtained upon increasing the hidden-unit density a.
this approach is comparable to that of standard (A) Accuracy for the 1D TFI model, at a few values of the field strength h and for an 80-spin chain with
ground-state QMC simulations (see supplemen- periodic boundary conditions (PBCs). Points below 10–8 are not shown to enhance readability. (B) Accuracy
tary materials). for the 1D AFH model, for an 80-spin chain with PBCs, compared with the Jastrow ansatz (horizontal
To validate our scheme, we consider the prob- dashed line). (C) Accuracy for the AFH model on a 10-by-10 square lattice with PBCs, compared with the
lem of finding the ground state of two prototyp- precision obtained by EPS [upper dashed line (35)] and PEPS [lower dashed line (36)]. For all cases

Downloaded from http://science.sciencemag.org/ on February 9, 2017

ical spin models, the transverse-field Ising (TFI) considered here, the NQS approach reaches MPS-grade accuracies in one dimension and systematically
model and the antiferromagnetic Heisenberg improves the best known variational states for 2D finite lattice systems.
(AFH) model. Their Hamiltonians are
X X 0.20
HTFI ¼ −h sxi − szi szj 1.0 A B
i ij
Jz →

|σiz σiz |(t)

h →
σx(t)

and 0.15
0.6
X Jz → /
HAFH ¼ sxi sxj þ syi syj þ szi szj h / →
ij
0.2 0.10
0 1 2 0 1 2 3
respectively, where s ; s ; and s are Pauli matrices.
x y z
t t
In the following, we consider the case of both
one- and two-dimensional (1D and 2D) lattices Fig. 4. Many-body unitary time evolution with NQS. NQS results (solid lines) for the time evolution
with periodic boundary conditions (PBCs). In induced by a quantum quench in the microscopic parameters of the models we study (the transverse
Fig. 2, we show the optimal network structure field h for the TFI model and the coupling constant Jz in the AFH model) are shown. (A) Time-dependent
of the ground states of the two spin models for transverse spin polarization in the TFI model, compared to exact results (dashed lines). (B) Time-
hidden-variable density a = 4 and with imposed dependent nearest-neighbors spin correlations in the AFH model, compared to exact numerical results
translational symmetries. We find that each fil- obtained with t-DMRG (dashed lines). All results refer to 1D chains representative of the thermodynamic
ter f ¼ ½1; …; a learns specific correlation features limit, with finite-size corrections smaller than the line widths.
emerging in the ground-state wave function. For
example, in the 2D case (Fig. 2, rightmost panels)
the neural network learns patterns correspond- error, which, for a small a = 4, attains the same energy upon increasing a, qualitatively similar
ing to antiferromagnetic correlations. The gen- high precision obtained for the TFI model at the to the 1D case. The increased difficulty of the
eral behavior of the NQS is completely analogous critical point. The accuracy of our model is sev- problem is reflected in a slower convergence. We
to that observed in convolutional neural networks, eral orders of magnitude higher than the spin- still obtain results at the level of existing state-
where different layers learn specific structures of Jastrow ansatz (dashed line in Fig. 3B). It is also of-the-art methods or better. In particular, with
the input data. interesting to compare the value of a with the a relatively small hidden-unit density (a ~ 4), we
In Fig. 3, we show the accuracy of the NQS, MPS bond dimension M needed to reach the already obtain results at the same level as the
quantified by the relative error on the ground- same level of accuracy. For example, on the AFH best-known variational ansatz for finite clusters
state energy Drel ¼ ðENQS ðaÞ−Eexact Þ=jEexact j, for model with PBCs, we find that with a standard [the entangled plaquette states (EPS) of (35)
several values of a and model parameters. In Fig. density matrix renormalization group (DMRG) and the projected entangled pair states (PEPS)
3A, we compare the variational NQS energies with implementation (33), we need M ~ 160 to reach of (36)]. Further increasing a then leads to a siz-
the exact result obtained by the fermionization of the accuracy NQS have at a = 4. This points able improvement and, consequently, yields the
the TFI model, on a 1D chain with PBCs. The toward a more compact representation of the best variational results reported to date for this
most notable result is that NQS achieve a con- many-body state in the NQS case, which fea- 2D model on finite lattices.
trollable and arbitrary accuracy that is compatible tures about three orders of magnitude fewer var-
with a power-law behavior in a. The hardest-to- iational parameters than the corresponding MPS Unitary dynamics
learn ground state is at the quantum critical point ansatz. NQS are not limited to ground-state problems
h = 1, where nonetheless a notable accuracy of We next studied the AFH model on a 2D but can be extended to the time-dependent
one part per million can be easily achieved with square lattice (for a comparison with QMC re- Schrödinger equation. For this purpose, we de-
a relatively modest density of hidden units. The sults, see Fig. 3C) (34). As expected from en- fine complex-valued and time-dependent network
same accuracy is obtained for the more complex tanglement considerations, the 2D case proves weights WðtÞ that, at each time t, are trained to
1D AFH model (Fig. 3B). In this case, we also ob- harder for the NQS. Nonetheless, we always best reproduce the quantum dynamics, in the sense
serve a systematic drop in the ground-state energy find a systematic improvement of the variational of the Dirac-Frenkel time-dependent variational

Carleo et al., Science 355, 602–606 (2017) 10 February 2017 3 of 4

R ES E A RC H | R E S EA R C H A R T I C LE

principle (37, 38). In this context, the variatio- Many paths for research can be envisaged in the 27. A. N. Kolmogorov, Dokl. Akad. Nauk SSSR 108, 179–182
nal residuals near future. For example, the most recent ad- (1961).
28. K. Hornik, Neural Netw. 4, 251–257 (1991).
vances in machine learning, like deep network 29. N. Le Roux, Y. Bengio, Neural Comput. 20, 1631–1649
RðẆ ðtÞÞ ¼ distð@t Y; −iHYÞ architectures and convolutional neural networks, (2008).
can constitute the basis of more advanced NQS 30. K. Sohn, H. Lee, “Learning invariant representations with
are the objective functions to be minimized as and therefore have the potential for increasing local transformations,” in Proceedings of the 29th International
Conference on Machine Learning, Edinburgh, Scotland, 26 June
a function of the time derivatives of the weights their expressive power. Furthermore, the extension to 1 July 2012 (Omnipress, 2012), pp. 1311–1318.
Ẇ ðtÞ (see supplementary materials). In the of our approach to treat quantum systems other 31. M. Norouzi, M. Ranjbar, G. Mori, “Stacks of convolutional
stochastic framework, this is achieved by a than interacting spins is, in principle, straight- restricted Boltzmann machines for shift-invariant feature
time-dependent VMC method (25, 26), which forward. In this respect, applications to answer learning,” in Proceedings of the IEEE Conference on Computer
Vision and Pattern Recognition, 20 to 25 June 2009 (IEEE, 2009),
samples jYM ðS; WðtÞÞj2 at each time and pro- the most challenging questions concerning inter- pp. 2735–2742.
vides the best stochastic estimate of the Ẇ ðtÞ acting fermions in two dimensions can already 32. S. Sorella, M. Casula, D. Rocca, J. Chem. Phys. 127, 014105
that minimizes R2(t), with a computational cost be anticipated. Finally, at variance with tensor (2007).
OðaN 2 Þ. Once the time derivatives are determined, network states, the NQS feature intrinsically 33. M. Dolfi et al., Comput. Phys. Commun. 185, 3430–3440
(2014).
these can be conveniently used to obtain the full nonlocal correlations, which can lead to substan- 34. A. W. Sandvik, Phys. Rev. B 56, 11678–11690 (1997).
time evolution after time integration. tially more compact representations of many- 35. F. Mezzacapo, N. Schuch, M. Boninsegni, J. I. Cirac, New J.
To demonstrate the effectiveness of the NQS body quantum states. A formal analysis of the Phys. 11, 083026 (2009).
in the dynamical context, we consider the unitary NQS entanglement properties might therefore 36. M. Lubasch, J. I. Cirac, M.-C. Bañuls, Phys. Rev. B 90, 064425
(2014).
dynamics induced by quantum quenches in the bring about substantially new concepts in quan-

Downloaded from http://science.sciencemag.org/ on February 9, 2017

37. P. A. M. Dirac, Math. Proc. Camb. Philos. Soc. 26, 376–385
coupling constants of our spin models. In the TFI tum information theory. (1930).
model, we induce nontrivial quantum dynam- 38. Y. I. Frenkel, Wave Mechanics: Advanced General Theory
ics by means of an instantaneous change in the (International Series of Monographs on Nuclear Energy:
RE FERENCES AND NOTES
Reactor Design Physics, The Clarendon Press, 1934).
transverse field: The system is initially prepared 1. D. Ceperley, B. Alder, Science 231, 555–560 (1986). 39. S. R. White, A. E. Feiguin, Phys. Rev. Lett. 93, 076401
in the ground state of the TFI model for some 2. W. M. C. Foulkes, L. Mitas, R. J. Needs, G. Rajagopal, Rev. Mod. (2004).
transverse field hi and then evolves under the Phys. 73, 33–83 (2001). 40. G. Vidal, Phys. Rev. Lett. 93, 040502 (2004).
3. J. Carlson et al., Rev. Mod. Phys. 87, 1067–1118 (2015). 41. A. J. Daley, C. Kollath, U. Schollwock, G. Vidal, J. Stat. Mech.
action of the TFI Hamiltonian with a transverse 4. S. R. White, Phys. Rev. Lett. 69, 2863–2866 (1992). 2004, P04005 (2004).
field hf ≠hi . We compare our results with the an- 5. S. Rommer, S. Ostlund, Phys. Rev. B 55, 2164–2181 42. B. Bauer et al., J. Stat. Mech. 2011, P05001 (2011).
alytical solution obtained from fermionization (1997).
of the TFI model for a 1D chain with PBCs. In 6. U. Schollwöck, Ann. Phys. 326, 96–192 (2011). AC KNOWLED GME NTS
7. R. Orús, Ann. Phys. 349, 117–158 (2014).
Fig. 4A, the exact results for the time-dependent 8. F. Verstraete, V. Murg, J. I. Cirac, Adv. Phys. 57, 143–224
We acknowledge discussions with F. Becca, J. F. Carrasquilla,
transverse spin polarization are compared to M. Dolfi, J. Osorio, D. Patané, and S. Sorella. The time-dependent
(2008).
MPS results have been obtained with the open-source
NQS with a = 4. In the AFH model (Fig. 4B), 9. K. H. Marti, B. Bauer, M. Reiher, M. Troyer, F. Verstraete, New J.
implementation available as a part of the Algorithms and Libraries
we study quantum quenches in the longitudinal Phys. 12, 103008 (2010).
for Physics Simulations (ALPS) project (33, 42). This work was
10. M. Troyer, U.-J. Wiese, Phys. Rev. Lett. 94, 170201 (2005).
coupling Jz and monitor the time evolution of 11. A. Polkovnikov, K. Sengupta, A. Silva, M. Vengalattore,
supported by the European Research Council (ERC) through ERC
the nearest-neighbors correlations. Our results Advanced Grant SIMCOFE, by the Swiss National Science
Rev. Mod. Phys. 83, 863–883 (2011).
Foundation through National Center of Competence
for the time evolution (with a = 4) are com- 12. J. Eisert, M. Friesdorf, C. Gogolin, Nat. Phys. 11, 124–130
in Research Quantum Science and Technology (QSIT), and by
pared with the numerically exact MPS dynam- (2015).
Microsoft Research. This paper is based on work supported
13. A. Montorsi, The Hubbard Model: A Collection of Reprints
ics (39–41) for a system with open boundaries (World Scientific, 1992).
in part by the Office of the Director of National Intelligence (ODNI),
(Fig. 4B). Intelligence Advanced Research Projects Activity (IARPA)
14. D. J. Thouless, The Quantum Mechanics of Many-Body Systems
via Massachusetts Institute of Technology Lincoln Laboratory Air
The high accuracy also obtained for the uni- (Reprint of the Academic Press, ed. 2, 1972).
Force contract no. FA8721-05-C-0002. The views and conclusions
tary dynamics further confirms that neural- 15. J. K. Freericks, B. K. Nikolić, O. Frieder, Int. J. Mod. Phys. B 28,
contained herein are those of the authors and should not
1430021 (2014).
network–based approaches can be successfully 16. G. E. Hinton, R. R. Salakhutdinov, Science 313, 504–507
be interpreted as necessarily representing the official policies
used to solve the quantum many-body problem, or endorsements, either expressed or implied, of ODNI, IARPA,
(2006).
or the U.S. government. The U.S. government is authorized to
not only for ground-state properties but also for 17. Y. LeCun, Y. Bengio, G. Hinton, Nature 521, 436–444
reproduce and distribute reprints for governmental purposes,
modeling the evolution induced by a complex (2015).
notwithstanding any copyright annotation thereon. The authors
18. D. Silver et al., Nature 529, 484–489 (2016).
set of excited quantum states. 19. S. S. Schoenholz, E. D. Cubuk, D. M. Sussman, E. Kaxiras,
agree to making the code used in this paper available upon
reasonable request.
A. J. Liu, Nat. Phys. 12, 469–471 (2016).
Outlook 20. J. Carrasquilla, R. G. Melko, https://arxiv.org/abs/1605.01735
Variational quantum states based on artificial (2016). SUPPLEMENTARY MATERIALS
21. L. Wang, Phys. Rev. B 94, 195105 (2016). www.sciencemag.org/content/355/6325/602/suppl/DC1
neural networks can be used to efficiently cap- 22. G. Torlai, R. G. Melko, Phys. Rev. B 94, 165134 (2016). Materials and Methods
ture the complexity of entangled many-body sys- 23. M. H. Amin, E. Andriyash, J. Rolfe, B. Kulchytskyy, R. Melko, Supplementary Text
tems in both one and two dimensions. Despite https://arxiv.org/abs/1601.02036 (2016). Figs. S1 and S2
the simplicity of the restricted Boltzmann ma- 24. W. L. McMillan, Phys. Rev. 138, A442–A451 (1965). References (43–45)
25. G. Carleo, F. Becca, M. Schiró, M. Fabrizio, Sci. Rep. 2, 243 Code and Data Files
chines used here, very accurate results for both (2012).
ground-state and dynamical properties of pro- 26. G. Carleo, F. Becca, L. Sanchez-Palencia, S. Sorella, M. Fabrizio, 7 June 2016; accepted 12 January 2017
totypical spin models can be readily obtained. Phys. Rev. A 89, 031602 (2014). 10.1126/science.aag2302

Carleo et al., Science 355, 602–606 (2017) 10 February 2017 4 of 4

 Solving the quantum many-body problem with artificial neural
networks
Giuseppe Carleo and Matthias Troyer (February 9, 2017)
Science 355 (6325), 602-606. [doi: 10.1126/science.aag2302]

Editor's Summary

Machine learning and quantum physics

Elucidating the behavior of quantum interacting systems of many particles remains one of the
biggest challenges in physics. Traditional numerical methods often work well, but some of the most
interesting problems leave them stumped. Carleo and Troyer harnessed the power of machine learning
to develop a variational approach to the quantum many-body problem (see the Perspective by Hush).
The method performed at least as well as state-of-the-art approaches, setting a benchmark for a

Downloaded from http://science.sciencemag.org/ on February 9, 2017

prototypical two-dimensional problem. With further development, it may well prove a valuable piece in
the quantum toolbox.
Science, this issue p. 602; see also p. 580

This copy is for your personal, non-commercial use only.

Article Tools Visit the online version of this article to access the personalization and
article tools:
http://science.sciencemag.org/content/355/6325/602

Permissions Obtain information about reproducing this article:

http://www.sciencemag.org/about/permissions.dtl

Science (print ISSN 0036-8075; online ISSN 1095-9203) is published weekly, except the last week
in December, by the American Association for the Advancement of Science, 1200 New York
Avenue NW, Washington, DC 20005. Copyright 2016 by the American Association for the
Advancement of Science; all rights reserved. The title Science is a registered trademark of AAAS.