Solving Schrodinger equations using physically constrained neural network

Kai-Fang Pu,1 Hanlin Li,1, ∗ Hong-Liang Lü,2 and Long-Gang Pang3, †

1
College of Science, Wuhan University of Science and Technology, Wuhan, Hubei 430065, China
2
HiSilicon Research Department, Huawei Technologies Co., Ltd., Shenzhen 518000, China
3
Key Laboratory of Quark and Lepton Physics (MOE) and Institute of Particle Physics,
Central China Normal University, Wuhan 430079, China
(Dated: March 8, 2023)
Deep neural network (DNN) and auto differentiation have been widely used in computational
physics to solve variational problems. When DNN is used to represent the wave function to solve
arXiv:2303.03934v1 [nucl-th] 7 Mar 2023

quantum many-body problems using variational optimization, various physical constraints have to
be injected into the neural network by construction, to increase the data and learning efficiency. We
build the unitary constraint to the variational wave function using
R x a monotonic neural network to
represent the Cumulative Distribution Function (CDF) F (x) = −∞ ψ ∗ ψdx0 . Using this constrained
neural network to represent the variational wave function, we solve Schrodinger equations using auto-
differentiation and stochastic gradient descent (SGD), by minimizing the violation of the trial wave
function ψ(x) to the Schrodinger equation. For several classical problems in quantum mechanics,
we obtain their ground state wave function and energy with very low errors. The method developed
in the present paper may pave a new way in solving nuclear many body problems in the future.
Keywords: Deep neural network; auto differentiation; variational problems; the Cumulative
Distribution Function; ground state wave function

PACS numbers:

I. Introduction can be helpful in finding low-dimensional manifolds in

high-dimensional space, which is crucial for the quantum
many-body problem, which suffers from the curse of di-
The universal approximation theorem of deep neural mensionality. The associated disadvantage is that it is
network (DNN) [1] makes it powerful in representing a at an early stage of development and its applicability to
variational function y = f (x, θ) with trainable param- computational physics has not been fully tested.
eters θ. In physics, this function can be used as so- With strong information encapsulation capability,
lutions of many different partial differential equations deep learning is proved to be a powerful tool in solv-
(PDEs) L̂f = 0, such as Maxwell equations for elec- ing quantum many-body problems [4–8]. The most typ-
tromagnetic field, Navier-stokes equations for fluid dy- ical application is to use DNN to represent the wave
namics, Schrodinger equations in quantum mechanics as function of quantum many-body states for many-electron
well as the Einstein field equations for gravity.The tradi- systems [9]. In subsequent developments, ANN applica-
tional way to solve this problem is to use physical mod- tions extended to prototypical spin lattice systems and
els. These models face great challenges in solving in- quantum systems in continuous space [10–12]. Recently
verse problems with complex geometric regions and high- machine learning has been used to deal with ab-initio
dimensional space. Unlike these models, the deep learn- problems [13–15]. The Feynman path integral [16] is an-
ing method developed in this paper provides a new di- other method for solve quantum state problems. Modern
rection to solve these problems. As the parameters of generative models can represent probability distribution
DNN are initialized with random numbers, the varia- with high computational efficiency. A Fourier-flow gener-
tional function f (x, θ) violates PDEs and the residuals ative model has been proposed to simulate the Feynman
δ = |L̂f | are usually the optimization objectives that can propagator and generate paths for quantum systems [17].
be minimized to desired precision. In this way, many And Ref [18] propose a Feynman path generator that can
physical problems [2] are naturally mapped into opti- estimate the Euclidean propagator and the ground state
mization problems [3] that can be solved using the mod- wave function with high accuracy.
ern deep learning libraries. PDEs usually have boundary and/or initial conditions.
The main advantages of machine learning are that In an early paper, these initial and boundary condi-
(1) it directly establishes the function mapping between tions are built into the neural network through construc-
input and output data, (2) ordinary differential equa- tion, the training objective is to minimize the resid-
tions (ODEs), PDEs can be transformed into variational ual δ alone. This method uses hard constraints such
problems that can be solved using optimization. It that f (x, θ) satisfies initial and boundary conditions au-
tomatically. It is thus quite data efficient. The re-
cent Physics Informed Neural Network [19–21] uses soft
constraints where the violation to initial and bound-
∗ Electronic address: [email protected] ary conditions are also added to the training objective
† Electronic address: [email protected] L = |L̂f | + β1 |δBC | + β2 |δIC |.
 2

Some variational functions should obey physical con-

straints. E.g., in solving the Maxwell equations, the mag- Z x
netic field represented by the DNN should be divergence F (x) = ψ ∗ (x0 )ψ(x0 )dx0 (1)
free. To take into this constraint, the paper ”Linearly −∞
constrained neural network” uses DNN to produce a vec- r
dF (x)
~ y, z, θ) whose curl ∇ × A
tor field A(x, ~ are divergence ψ(x) = (2)
dx
free [22]. It is thus also possible to construct a scalar
field φ(x, y, z, θ) whose gradients (∂x φ, ∂y φ, ∂z φ) are curl where F (x) is the CDF represented by a neural network
free. Actually, a general method has been developed that is monotonic by construction.We choose to constrain
to construct neural networks with linear constraints. In the weights to make the algorithm data efficient. The
solving the many body Schrodinger equations, the many derivative dF/dx is calculated using auto differentiation
fermion wave function should be anti-symmetric. Fer- (auto-diff) [35], that is provided by the deep learning li-
miNet, PauliNet uses Slater determinant to construct braries automatically, in analytical precision. There are
DNNs that are anti-symmetric. [23, 24] In DFT [25–27] two advantages using CDF. First, the wave function ex-
and Molecular dynamics [28], the local chemical environ- tracted from CDF automatically satisfies the normaliza-
ment usually have translational or rotational symmetry tion condition. So there is no numerical integral in the
that is considered using gauge equivalent neural network whole calculation. Second, the values of CDF is in be-
[29]. In Lattice Gauge field theory [30], Gauge equivari- tween (0, 1), whose range is much smaller than PDF,
ant normalizing flows are employed to sample field con- making the neural network much easier to train under
figurations [31]. the same learning rate. In practice, our training epochs
In the present paper, we use a monotonic neural net- are far fewer than the previous algorithm. And because
Rwork
x
to represent the cumulative distribution function we eliminate all the integrals, the computation of each
−∞
f (x0 )dx0 , whose first order derivative is the prob- epoch is also less than previous algorithms. So our algo-
ability density f (x) = ψ ∗ (x)ψ(x) that gives the ground rithm can achieve higher accuracy with less computation.
state wave function.The present paper demonstrates that We use feed forward neural network, or simply multi-
neural network with physical constraints can be used as layer percepton to represent the CDF. The input of the
efficient trial wave functions of Schrodinger equations. neural network is the n-dimensional spatial coordinates
Auto-diff helps to compute the required derivatives of x, the first layer of the DNN consists m = 32 hidden neu-
the trial function with respect to the input variables. In rons whose values are calculated by h1 = σ(xW1 + b1 ),
this way, optimizing the violation of the trial function where W1 is the weight matrix with n×m elements and b1
to PDEs solves PDEs to high accuracy. Compared to is the bias vector with m values. The σ is the activation
previous methods, our calculation does not need to cal- function which brings neural network non-linear repre-
culate any numerical integrals in the whole calculation sentation ability. To increase the representation power,
and the unitary constraint we impose on the variational the values of neurons in the first hidden layer are feed
wave function increases the data learning efficiency. The forward to the second hidden layer with similar opera-
improved algorithm greatly reduces the amount of com- tions h2 = σ(h1 W2 + b2 ). One can stack multiple hidden
putation required to solve the same Schrodinger equa- layers with one output neuron in the last layer to repre-
tion. These advantages make our method more suitable sent the value of CDF function. The whole neural net-
for dealing with many-body state, which require a huge work can thus be thought of as one variational function
amount of computation. F (x, θ) with θ all the trainable parameters in the neural
network.
To make sure that F (x, θ) is monotonic, we add a non-
negative constraint to the weights Wi of the neural net-
II. Methods work. At the same time, the activation function should
also be monotonic. In principle, sigmoid, tanh as well
The traditional variational method for quantum me- as leaky relu can all be used to construct this monotonic
chanics [32, 33] usually use given function with unknown neural network. In practice, we use sigmoid activation
parameters as variational function, e.g., e−αr with α one function whose derivatives are also continuous. This is
unknown parameter. Different from the previous Vari- important when the second order derivatives are required
ational Artificial Neural Network (VANN) applications in auto-diff. E.g., if one uses relu activatition function,
[24, 34], we do not use DNN to represent the wave func- the second order derivatives of the output of the neural
tion directly, instead, we use DNN to represent the CDF, network to the input equal 0. The last layer also uses
which is the integration of the probability density func- sigmoid activation function to make sure that the output
tion on the spatial coordinate.The training objective is range is (0, 1).
thus to minimize the violation of the wave function ψ(x) The training objective is to find the ground state en-
represented by the neural network to the Schrodinger ergy E0 and its corresponding wave function ψ0 by min-
equation. Its relationship with the wave function are imizing the violation of the wave function ψ(x) to the
shown as follows, following Schrodinger equation,
 3

problems can be regarded as many independent harmonic

oscillators whose potential in the Hamiltonian can be
H|ψi = E0 |ψi (3) written as,
2
where H = − 2m h̄
∇2 + V (x) is the Hamiltonian opera- 1
V = mω 2 x2 (5)
tor and E0 represents its smallest eigenvalue. The loss 2
function is thus set to be, where m is the mass of the oscillator, ω is its angular fre-
L(θ) = |(H − E0 )|ψi + |F (xmin )| + |F (xmax ) − 1| (4) quency and x is its deviation from equilibrium position.
The second problem is to solve Schrodinger equation
where θ represent all the trainable parameters in the with Woods-Saxon potential [38] that is widely used in
monotonic neural network, the ∇2 |ψi is computed by the nuclear physics to represent the charge distributions of
neural network through auto-diff, E0 is another trainable nucleus,
parameter initialized with constant number 0.0. Two −1
additional loss term are added to take into account the V = |x|−R0
(6)
boundary condition of the CDF. We use it to limit the 1+e a0

range of values of the CDF, which ensures that the wave-

function satisfies the normalization condition. In previ- Where a0 is related to the thickness of the surface layer,
ous Variational Artificial Neural Network (VANN) appli- in which the potential drops from the outside to the inside
cations, this term was written as < ψ|H|ψ > / < ψ|ψ >. of the nucleus and R0 is the average radius of the nucleus
We replace the numerical integration of the denominator at which the average interaction occurs.
with soft constraints, which simplifies the calculation. The third potential is an infinitely high potential well
Before optimization, the weight values of the neural with a width of 2l,
network parameters are usually initialized randomly or
through Xavier scheme [36]. In our problem, we observe 
∞, |x| > l
that the scheme of parameter initialization has little in- V = (7)
0, |x| ≤ l
fluence on the training process and the result of varia-
tional optimization. For the sake of brevity, the parameters in Hamiltonian
We try to eliminate all numerical integrals in the whole use the following values,
calculation. Because in neural network calculation, the
differential is easier to calculate than integral. To calcu- h̄ = m = ω = 1, R0 = 6.2, a0 = 0.1, l = 4. (8)
late the integral, we have to use numerical approxima-
tion methods such as Monte Carlo sampling, which will Different from previous studies that solve Schrodinger
certainly increase the amount of computation and may equations using supervised learning, our method is close
affect the accuracy. In practice, we found that we had to to unsupervised learning where both E0 and ψ0 are
subtract the energy term if we didn’t want to do the inte- learned through optimization. The input to the neural
gral.To find the ground state energy and wave function, network is a list of shuffled coordinates sampled from the
we add another loss term e0.001E0 . The basic logic is to domain. Using these coordinates, we compute the loss
decrease E0 during the optimization of the overall loss. functions and minimize the violation of E0 and ψ0 to the
The function form of this loss term is designed to produce Schrodinger equation, as well as e0.001E0 . In principle, we
proper negative gradients at different stages of training. can use markov chain monte carlo (MCMC) method [39]
First, the gradients should be large to make it converge to sample coordinates with the learned wave function, or
fast when E0 is much larger than the ground state energy. use active learning to generate coordinates that violate
Second, the gradients should be small enough to avoid in- Schrodinger equations more with the currently learned
terfering with the optimization of other parts of overall network, to speed up the training process. In practice,
loss when E0 is close enough to the true value. Last but for these simple problems, the wave function are usually
not least, this loss term must monotonously increase with very close to the exact wave function after training the
E0 throughout the definition domain. In principle, E0 DNN for 2000 iterations. We generally train 10,000 iter-
can be smaller than the analytical ground state energy, ations with a very small learning rate for the last 1000
however, in that case the residual of Shrodinger equation iterations.
increases faster than this term because of the coeficient We use tensorflow [40] to construct the DNN, to com-
0.001. After trained, the value of E0 approaches the ex- pute the auto-diff dF/dx as well as ∇2 ψ and to update
act value of the ground state energy. the network parameters. We use Adam algorithm [41]
We test the performance of the DNN Schrodinger equa- that add momentum mechanism and adaptive learning
tion solver on three classical quantum mechanical prob- rate to thePm simple stochastic gradient descent θn+1 =
1 ∂li
lems. The first problem is the harmonic oscillator prob- θn − lr m i=1 ∂θ . The relevant parameters in Adam
lem [37]. Harmonic oscillator is used to approximate algorithm are set to β1 = 0.9, β2 = 0.999,  = 10−7 .
molecular vibration, lattice vibration, radiation field vi- To speed up the training process, we use learning rate
bration and so on, around the steady point. All of these scheduler to adjust the learning rate and make it vary
 4

between 10−2 −10−5 . A large learning rate at early stage

makes the function converge faster at the beginning, and
a small learning rate at late time makes the training pro-
cess smooth.
To quantify the difference between the true wave func-
tion ψtrue and the wave function learned by the DNN
ψDNN , we introduce partial-wave fidelity K in the follow-
ing,
< ψtrue |ψDNN >2
K= (9)
< ψtrue |ψtrue >< ψDNN |ψDNN >
The closer the K is to one, the closer the result of DNN FIG. 1: The cumulative distribution equation as a function
of position in harmonic oscillator problem
is to the exact wave function.

III. Results

For the harmonic oscillator potential used in the

present paper, the analytical ground state energy is
0.5h̄ω. After 1500 iterations of training, the ground state
energy from DNN is E0 = 0.50038h̄ω, whose relative er-
ror is within 0.06%. In the last stage of 10000 iterations,
the error can be controlled below 0.002%.
The partial-wave fidelity K is approximately 0.997993
using 3 hidden layers with 32 units (neurons) per layers.
To study the influence of the number of variational pa-
rameters on the training results, we computed K using
different numbers of hidden layers and numbers of units
per layer. The results are shown below.

Nlayer
1 2 3 4
Nunit
4 0.9995717 0.9999767 0.9999705 0.9999618
8 0.9999416 0.9999797 0.9999910 0.9999932
16 0.9999861 0.9999923 0.9999936 0.9999967
32 0.9999789 0.9999909 0.9999896 0.9999922
64 0.9999744 0.9999746 0.9999903 0.9999941

TABLE I: The fidelities of VANN result,the first row rep-

resents the number of hidden layers, and the first column
represents the number of units in each layer

As shown in Table.I, the highest fidelity happens us-

ing 4 hidden layers with 16 hidden neurons per layer,
with K = 0.9999967 . For this simple problem, the DNN
achieve a very low error in fidelity even with only two
hidden layers and four units per layer. The performance FIG. 2: The ground state wave function(top panel),first
increases with the the number of variational parameters, derivative(central) and the second derivative(bottom) as a
to approach the exact wave function. However, the per- function of the position in harmonic oscillator problem.
formance of DNN saturate or even decrease if there are
too many variational parameters.
To further visualize the difference between the result As shown in Fig. 1, the difference between DNN results
of DNN and the exact solution, we compare the CDF in and the ground truth is within the error range of 0.0001.
FIG.1. p The error range of the ground state wave function can be
Using Eq. 2, we compute ψ(x) = dF/dx and com- controlled within 0.0002 as shown in Fig. 2. In addition,
pare the wave function ψ(x), its first and second deriva- the accuracy of the learned first and second order deriva-
d2 ψ
tives dψ
dx and dx2 with the analytical results. This pro- tives through variational optimization are also very high,
vides a detailed comparison that shows the power of the which means that this method is not only accurate, but
variational function represented by the DNN. also captures the true physics in stead of finding an ap-
 5

proximation function for the grond state wave function. from zero to infinity at the boundary. We think that this
is due to the fact that the soft constraint cannot handle
the infinite potential energy at the boundary well, so the
probability density at the boundary is not 0.

IV. Conclusions

In the present paper, we use physics based neural net-

work to solve Schrodinger equations numerically. We de-
signed a monotonic neural network to represent the CDF
of the ground state wave function. In this way, the wave
function represented by the DNN satisfy the normaliza-
FIG. 3: The ground state wave function in the Woods-Saxon tion condition by design. The variational optimization
potential energy is reduced to a optimization problem by minimizing the
violation of the trial wave function and trial ground state
energy E0 to Schrodinger equations. The method is used
to solve Schrodinger equations with 3 different potentials,
the harmonic oscillator, the Woods-Saxon potential and
the infinitely high potential well, all with small relative
error.
Compared to traditional variational methods in solving
quantum mechanical problems, the trial wave function
represented by DNN do not have fixed function forms
before training. On the other hand, the training objec-
tive is different from the traditional E0 = hψ|H|ψi
hψ|ψi where
numerical integration is required for both the numera-
tor and the denominator. In our case, the objective is
FIG. 4: The ground state wave function in infinitely high to minimize the violation to the Schrodinger equation,
potential well on sampled spatial coordinates. As the neural network
is constrained, the trial wave function is normalized by
The same network is used to to solve Schrodinger equa- construction. Our method is also different from the pre-
tions with Woods-Saxon potential and the infinitely high vious Schrodinger equation solver using supervised learn-
potential well. No modification is made to hyperparam- ing, where ground state energies from numerical solutions
eters other than the potential part in the Hamiltonian. are needed to train the neural network. In another DNN
The comparisons between the learned ground state wave Schrodinger solver [32, 33], the initial values of the net-
functions and the true values are shown in Fig. 3 and work parameters greatly affect the optimization results.
Fig. 4. The result shows that the ground state wave func- To avoid strong fluctuations, they provide a trial wave
tions obtained from the DNN CDF are also in excellent function whose form is close to the exact solution. The
agreement with the exact solution. The error range of disadvantage of the previous algorithms is that it can
the Woods-Saxon potential’s ground state wave function only solve the problem that the form of exact solution of
can be controlled within 0.0002. As shown in Fig.4, the the equation is known. Our algorithm can directly ignore
performance of the network on infinitely high potential the pre-training process, so we do not need to know any
well is relatively poor, whose range of error is expanded information of the exact solution before training, which
to 0.02, which is much worse than harmonic oscillator po- is more universal and provides the possibility to solve
tential and Woods-Saxon potential. This is also reflected problems that have never been dealt with before. In ad-
in the calculation of ground state energy and the partial- dition, we observe that our DNN can approximate the
wave fidelity K. The ground state energy calculated by ground state wave function with fewer trainable param-
DNN for the Woods-Saxon potential problem is -0.97382, eters. And the physical constraints constructed in the
also within an error of 0.002% to the exact result -0.97385 neural network make the current method quite data effi-
and K = 0.999964, similar to the harmonic oscillator po- cient. So that we can achieve higher accuracy with less
tential. While for the potential well problem, the DNN computation.
E0 = 0.07787, whose relative error to the accurate result The current method can be improved in several ways.
0.07710 is about 1% and K = 0.9977475 also less than First, the CDF works for wave functions in high di-
the average level of the first two potentials. It reminds us mensional space as long as n-dim spatial coordinates
that this DNN might not perform good at dealing with are flattened. Second, the spatial coordinates used for
potentials with discontinuities, like infinitely high poten- training can be sampled using the learned wave function
tial well, whose potential energy discontinuously change or through active learning, to increase the training effi-
 6

ciency. Third, the anti-symmetric constraints of the wave No.12075098), the Natural Science Foundation of Hubei
function should be considered for many fermion systems. Province (Grant No.2019CFB563), the Hubei Province
Although further efforts have to be done to improve the Department of Education (Grant No. D20201108), Hubei
current method, it shows good properties in solving clas- Province Department of Science and Technology (Grant
sical quantum mechanical problems. The next step is No. 2021BLB171). LG Pang and KF Pu also acknowl-
to solve the ground state energy and wave functions of edge the support provided by Huawei Technologies Co.,
Deuteron. It also paves a new way in solving many nu- Ltd. Computations are performed at the NSC3 super
cleon problems. cluster at CCNU and High-Performance Computing Cen-
ter of Wuhan University of Science and Technology.

Acknowledgments

This work is supported by the National Natural

Science Foundation of China (Grants No. 12035006,

[1] G. Cybenko, Mathematics of Control, Signals, and Sys- [17] S. Chen, O. V. Savchuk, S. Zheng, B. Chen, H. Stoecker,
tems (MCSS) 2, 303 (1989), ISSN 0932-4194, URL http: L. Wang, and K. Zhou, Physical Review D (2022).
//dx.doi.org/10.1007/BF02551274. [18] Y. Che, C. Gneiting, and F. Nori, ArXiv: 2202.02750
[2] A. Boehnlein, M. Diefenthaler, N. Sato, M. Schram, (2022).
V. Ziegler, C. Fanelli, M. Hjorth-Jensen, T. Horn, [19] M. Raissi, P. Perdikaris, and G. E. Karniadakis, ArXiv:
M. P. Kuchera, D. Lee, et al., Rev. Mod. Phys. 94, 1711.10561 (2017).
031003 (2022), URL https://link.aps.org/doi/10. [20] M. Raissi, P. Perdikaris, and G. E. Karniadakis, ArXiv:
1103/RevModPhys.94.031003. 1711.10566 (2017).
[3] D. J. MacKay and D. J. Mac Kay, Information theory, [21] M. Raissi, P. Perdikaris, and G. E. Karniadakis, Journal
inference and learning algorithms (Cambridge university of Computational Physics 378, 686 (2019).
press, 2003). [22] J. N. Hendriks, C. Jidling, A. G. Wills, and T. B. Schön,
[4] P. Mehta, M. Bukov, C.-H. Wang, A. G. R. Day, ArXiv: 2002.01600 (2020).
C. Richardson, C. K. Fisher, and D. J. Schwab, Physics [23] D. Pfau, J. S. Spencer, A. G. D. G. Matthews,
reports 810, 1 (2018). and W. M. C. Foulkes, Phys. Rev. Res. 2, 033429
[5] E. M. Nordhagen, J. M. Kim, B. Fore, A. Lovato, and (2020), URL https://link.aps.org/doi/10.1103/
M. Hjorth-Jensen arXiv:2210.00365 (2022). PhysRevResearch.2.033429.
[6] B. R. Barrett, P. Navratil, and J. P. Vary, Prog. Part. [24] J. Hermann, Z. Schätzle, and F. Noé, Nature Chemistry
Nucl. Phys. 69, 131 (2013). pp. 1–7 (2019).
[7] G. Torlai, G. Mazzola, J. Carrasquilla, M. Troyer, [25] P. Hohenberg and W. Kohn, Phys. Rev. 136,
R. Melko, and G. Carleo, Nature Physics 14, 447 (2018). B864 (1964), URL https://link.aps.org/doi/10.
[8] C. Adams, G. Carleo, A. Lovato, and N. Rocco, Phys. 1103/PhysRev.136.B864.
Rev. Lett. 127, 022502 (2021), URL https://link.aps. [26] W. Kohn and L. J. Sham, Phys. Rev. 140,
org/doi/10.1103/PhysRevLett.127.022502. A1133 (1965), URL https://link.aps.org/doi/10.
[9] G. Carleo, I. Cirac, K. Cranmer, L. Daudet, M. Schuld, 1103/PhysRev.140.A1133.
N. Tishby, L. Vogt-Maranto, and L. Zdeborová, Rev. [27] J. B. Condon, Surface area and porosity determinations
Mod. Phys. 91, 045002 (2019), URL https://link.aps. by physisorption: Measurement, classical theories and
org/doi/10.1103/RevModPhys.91.045002. quantum theory (Elsevier, 2019).
[10] M. Ruggeri, S. Moroni, and M. Holzmann, Phys. Rev. [28] M. Badar, S. Shamsi, J. Ahmed, and A. Alam, Molecular
Lett. 120, 205302 (2018), URL https://link.aps.org/ Dynamics Simulations: Concept, Methods, and Applica-
doi/10.1103/PhysRevLett.120.205302. tions (2020), ISBN 978-3-030-94650-0.
[11] J. Han, L. Zhang, and E. Weinan, J. Comput. Phys. 399 [29] D. Luo, G. Carleo, B. K. Clark, and J. Stokes, Phys. Rev.
(2018). Lett. 127, 276402 (2021), URL https://link.aps.org/
[12] S. Shi, K. Zhou, J. Zhao, S. Mukherjee, and P. Zhuang, doi/10.1103/PhysRevLett.127.276402.
Phys. Rev. D 105, 014017 (2022), URL https://link. [30] M. Creutz and K. J. M. Moriarty, Journal of Physics G:
aps.org/doi/10.1103/PhysRevD.105.014017. Nuclear Physics 9 (1983).
[13] K. Choo, A. Mezzacapo, and G. Carleo, Nature Commu- [31] R. Abbott, M. S. Albergo, A. Botev, D. Boyda, K. Cran-
nications 11 (2019). mer, D. C. Hackett, G. Kanwar, A. G. de G. Matthews,
[14] M. Scherbela, R. Reisenhofer, L. Gerard, P. Marque- S. Racanière, A. Razavi, et al., Proceedings of The 39th
tand, and P. Grohs, Nature Computational Science 2, International Symposium on Lattice Field Theory ???
331 (2021). PoS(LATTICE2022) (2022).
[15] Y. Yang, P. Zhao, arXiv:2211.13998,(2022). [32] J. Keeble and A. Rios, Physics Letters B 809, 135743
[16] R. P. Feynman, Rev. Mod. Phys. 20, 367 (1948), URL (2020).
https://link.aps.org/doi/10.1103/RevModPhys.20. [33] H. Saito, Journal of the Physical Society of Japan 87,
367. 074002 (2018).
 7

[34] G. Carleo and M. Troyer, Science 355, 602 (2017). [38] M. Capak and B. Gönül, Modern Physics Letters A 31,
[35] A. Paszke, S. Gross, S. Chintala, G. Chanan, E. Yang, 1650134 (2016).
Z. DeVito, Z. Lin, A. Desmaison, L. Antiga, and A. Lerer [39] S. Marsland, Markov Chain Monte Carlo Methods (Ma-
(2017). chine Learning,2014).
[36] X. Glorot and Y. Bengio, in Proceedings of the thir- [40] M. Abadi, A. Agarwal, P. Barham, E. Brevdo, Z. Chen,
teenth international conference on artificial intelligence C. Citro, G. S. Corrado, A. Davis, J. Dean, M. Devin,
and statistics (JMLR Workshop and Conference Pro- et al., arXiv preprint arXiv:1603.04467 (2016).
ceedings, 2010), pp. 249–256. [41] D. P. Kingma and J. Ba, arXiv preprint arXiv:1412.6980
[37] S. Garrett, The Simple Harmonic Oscillator (2020), pp. (2014).
59–131, ISBN 978-3-030-44786-1.