Neural Network Gauge Field Transformation for 4D SU(3)

gauge fields
arXiv:2405.19692v1 [hep-lat] 30 May 2024

Xiao-Yong Jin𝑎,∗
𝑎 Computational Science Division, Argonne National Laboratory,
Lemont, IL 60439, USA
E-mail: [email protected]

We construct neural networks that work for any Lie group and maintain gauge covariance, enabling
smooth, invertible gauge field transformations. We implement these transformations for 4D SU(3)
lattice gauge fields and explore their use in HMC. We focus on developing loss functions and
optimizing the transformations. We show the effects on HMC’s molecular dynamics and discuss
the scalability of the approach.

The 40th International Symposium on Lattice Field Theory (Lattice 2023)

July 31st - August 4th, 2023
Fermi National Accelerator Laboratory

∗ Speaker

© Copyright owned by the author(s) under the terms of the Creative Commons
Attribution-NonCommercial-NoDerivatives 4.0 International License (CC BY-NC-ND 4.0). https://pos.sissa.it/
Neural Network Gauge Field Transformation for 4D SU(3) gauge fields Xiao-Yong Jin

1. Introduction

Neural networks can approximate arbitrary functions [1–4]. When combined with gauge symme-
tries [5, 6] and adapted to lattice fields [7–12], they offer new tools for developing algorithms [13–20]
in lattice QCD. One key application is improving the efficiency of the Hybrid Monte Carlo (HMC)
algorithm [21].
HMC uses Hamiltonian dynamics to generate proposal configurations for a Markov Chain,
sampling the configuration space according to a known unnormalized probability density function.
The Hamiltonian dynamics, however, can struggle with potential barriers due to limited energy or
may linger in lengthy potential valleys. Lattice QCD systems, with their high degrees of freedom and
separation of long and short distance modes, exhibit critical slowing down [22] near the continuum
limit, with the cost of generating independent gauge field configurations also increasing as the force
in the Hamiltonian dynamics requires finer numerical integration steps. This conference offers two
reviews [23, 24] on improvements to HMC algorithms.
This work focuses on creating smooth and invertible gauge field transformations with tunable
parameters, incorporating neural networks while ensuring gauge covariance. Generalizing prior
work on the perturbative Wilson flow [25] and extending a study in 2D U(1) gauge field [10], we
apply tuned transformations in HMC with 4D SU(3) gauge fields.
We first describe the gauge field transformation, then introduce the neural networks, and present
results in optimizing and applying the transformation in molecular dynamics. We discuss the
scalability and transferability of our approach across different lattice volumes and gauge couplings.

2. Gauge field transformation

We introduce a continuously differentiable bijective map F with 𝑈 = F (𝑉) on the gauge fields
𝑉 ↦→ 𝑈, and apply a change of variables in the path integral for an observable O,
∫ ∫
1 1
D[𝑈]O (𝑈)𝑒 −𝑆 (𝑈) = D[𝑉]O F (𝑉) 𝑒 −𝑆FT (𝑉 ) ,


⟨O⟩ = (1)
Z Z
with the effective action after the field transformation,


𝑆FT (𝑉) = 𝑆 F (𝑉) − ln F∗ (𝑉) , (2)

and the Jacobian of the transformation,

𝜕F (𝑉)
F∗ (𝑉) = . (3)
𝜕𝑉
We construct the gauge field transformation with
" #
∑︁
𝑈 𝑥, 𝜇 = exp 𝜖 𝑥, 𝜇,𝑙 𝜕𝑥, 𝜇 𝑊 𝑥, 𝜇,𝑙 𝑉𝑥, 𝜇 , (4)
𝑙

where 𝑉𝑥, 𝜇 denotes the gauge link from lattice site 𝑥 to 𝑥 + 𝜇, ˆ in the direction 𝜇. The Wilson loop,
𝑊 𝑥, 𝜇,𝑙 , is an ordered product of gauge links along the loop labeled 𝑙 that goes through the link at
𝑥, 𝜇. The group derivative, 𝜕𝑥, 𝜇 𝑊 𝑥, 𝜇,𝑙 , is with respect to the gauge link, 𝑉𝑥, 𝜇 .

2
Neural Network Gauge Field Transformation for 4D SU(3) gauge fields Xiao-Yong Jin

(a) (b) (c)

Figure 1: Wilson loops used in smearing update to the red link in a 3D lattice. From left to right, (a) the links
in black used to update the red link, (b) four plaquette loops, (c) two 6-link chair shaped loops with one side
perpendicular to, and at the right side of, the red link.

Generalizing the stout smear [26] or the Wilson flow [25], we optionally make the coefficients
depend on Wilson loops (𝑋, 𝑌 , . . . ) independent of 𝑉𝑥, 𝜇 ,
2
tan−1 N𝑙 (𝑋, 𝑌 , . . .) .
 
𝜖 𝑥, 𝜇,𝑙 = 𝑐 𝑙 (5)
𝜋
The additional application of tan−1 and the multiplication of scalar coefficients 𝑐 𝑙 serve to constrain
the possible values of 𝜖, ensuring the resulting Jacobian determinant is positive definite. The
independence of the Wilson loops (𝑋, 𝑌 , . . . ) from 𝑉𝑥, 𝜇 simplifies the computation of the Jacobian
determinant and results in the constraint [25]
3
𝑐𝑙 = , (6)
4𝑁𝑙
for 𝑁𝑙 the number of distinct Wilson loops 𝑙. Naturally, N𝑙 can be any neural network.
In practice, we update only a subset of the gauge links at a time, a single direction out of four
space-time directions, and on either even or odd lattice sites lexicographically. For each subset, we
pick the derivatives of the Wilson loops, 𝜕𝑥, 𝜇 𝑊𝑙 , such that the only link to be updated in 𝑊𝑙 is 𝑥, 𝜇.
This ensures that the Jacobian matrix of the transformation is diagonal. Updating the whole lattice
requires eight separate subset updates.
Figure 1 shows the Wilson loops used for taking derivatives, in equation (4), when updating a
single gauge link, in our implementation. While the figure depicts a 3D lattice, in 4D space-time,
there are 6 plaquette loops and 48 chair loops, which go through the updating link marked red in
the figure. The derivatives of those loops with respect to the red link generate the basis, the linear
combination of which forms the 𝑠𝑢(3) algebra space in the exponential map in equation (4).

3. Neural network architectures

Stacking layers of individual transformations as in equation (4) forms a deep residual neural
network [9] generalized for gauge fields. Further generalizing it with standard neural networks,

3
Neural Network Gauge Field Transformation for 4D SU(3) gauge fields Xiao-Yong Jin

(a) (b) (c) (d)

Figure 2: Input to the neural network that computes smearing coefficients for updating the red link in a 3D
lattice. From left to right, (a) the links in black used to compute Wilson loops as input to a neural network, (b)
two 6-link rectangle loops parallel to the red link, (c) four 6-link rectangle loops perpendicular to the red link
on one side, (d) four plaquette perpendicular to the red link on one side.

we employ basic neural network layers that sequentially apply to traced Wilson loops and produce
localized smearing coefficients 𝜖 𝑥, 𝜇,𝑙 as introduced in equation (5). As mentioned, to simplify the
computation of the Jacobian matrix, the traced Wilson loops used at each layer are independent of
the links to be updated at that layer with equation (4).
To limit implementation complexity, we use only plaquette and 2 × 1 rectangular traced Wilson
loops for computing local smearing coefficients, as shown in figure 2 for updating the red link
(showing a 3D lattice as an example). In a 4D space-time lattice gauge field, for each link to be
updated, we use four plaquettes in each of the three planes parallel to and on each side of the updating
link, and the same amount of the 2 × 1 rectangular loops in those planes, and three 2 × 1 rectangular
loops that are parallel to the updating link. These amount to 51 distinct loops that are locally close to
the updating link. To directly use available neural network implementations from TensorFlow [27],
we compute the traces of the ordered product, and the traces of their squared and cubic powers, of
the gauge links along the loops, and then use the real and imaginary parts of the complex numbers
as the input to the traditional neural networks, resulting in six real floating point numbers per Wilson
loop, and a total of 306 real numbers per lattice site.

𝑌𝑥, 𝜇,𝑙 = {Re 𝑦 𝑥, 𝜇,𝑙,𝑖 , Im 𝑦 𝑥, 𝜇,𝑙,𝑖 }𝑖=1,2,3 (7)

𝑖
𝑦𝑖 = tr 𝑊 𝑥, 𝜇,𝑙 (8)

Instead of directly calling 4D convolutional neural networks, we implement symmetrical shift,

which shifts all the traces of Wilson loops of a fixed distance to the lattice link to be updated, and
then call dense neural networks,
𝑌𝑥,out = 𝜎(𝑊𝑌𝑥,in + 𝑏), (9)
where the weight matrix 𝑊 and bias vector 𝑏 are shared among the lattice sites inside a particular
layer. We choose Swish [28] as the activation function, 𝜎, for its continuous derivative. We omit the

4
Neural Network Gauge Field Transformation for 4D SU(3) gauge fields Xiao-Yong Jin

index 𝜇 here, since we always update links in a single direction per neural network layer.
We also introduce a normalization layer similar to typical layer-normalization in computer
vision, but we use site-local scaling coefficients (𝛾𝑖 ) and shift (𝛽𝑖 ), which have the same dimension
as the vector at each site and have the same values across the lattice.
1 ∑︁
𝑚= 𝑌𝑥,𝑖,in , (10)
𝐷𝑉/2 𝑥,𝑖
𝑌˜𝑥,𝑖 = 𝑌𝑥,𝑖,in − 𝑚 (11)
2 1 ˜2
𝜎𝑚 = 𝑌 (12)
𝐷𝑉/2 𝑥,𝑖
𝑌𝑥,𝑖,out = 𝛾𝑖𝑌˜𝑥,𝑖 /𝜎𝑚 + 𝛽𝑖 (13)

where 𝐷 is the vector dimension per site, 𝑉 is the lattice volume, and the summation goes through
the lattice (single parity) and the vector index.
Additionally, we implemented self-attention, feedforward, and residue networks, which operate
on numerical vectors on a 4D lattice. However, neural network layers using these operations exceed
the available compute and memory for reasonably large lattice volumes.
As we update the gauge links in a single direction on either even or odd lattice sites in a layer,
this computation creates a vector of real floating-point numbers on each lattice site of a given
parity, even or odd. Each operation—symmetric shift, site-local dense network, normalization,
and etc.—applies to the lattice field of vectors on a single parity and returns a vector field of the
same lattice size with optionally different degrees of freedom per site. Thus, we can compose many
layers of sequential operations that create the local coefficients used for gauge field transformation in
the smearing update as in equation (4). We sequentially apply transformation layers with different
weights to gauge link subsets in a single direction and parity. This transforms the whole lattice gauge
field, yielding a continuously differentiable, bijective deep neural network map of the fields.
Our code for HMC with neural network field transformation is available online [29].

4. Test result

We use the DBW2 action with 𝑐 1 = −1.4088, 𝛽 = 0.7796, corresponding to 𝑎/𝑟 0 ≃ 0.4 and
𝑎 = 0.2000(20) as a baseline following Ref [30, 31]. We optimize the map so that the gauge force
of the effective action matches an action with a stronger gauge coupling,

′


𝑆FT (𝑉; 𝛽 = 0.7796) = 𝜕𝑥, 𝜇 𝑆 F (𝑉; 𝛽 = 0.7796) − ln F∗ (𝑉) (14)
∼ 𝑆 ′ (𝑉; 𝛽 = 𝛽T ). (15)

As 𝛽 is a multiplicative constant in the gauge action, matching forces effectively seeks a transformation
that reduces the gauge force per link by a constant factor.
Denote the difference in force computed from the transformed action and the original action
with different couplings as,

Δ 𝑥, 𝜇,𝑐 = 𝜕𝑥, 𝜇,𝑐 𝑆FT (𝑉; 𝛽) − 𝜕𝑥, 𝜇,𝑐 𝑆(𝑉; 𝛽T ), (16)

5
Neural Network Gauge Field Transformation for 4D SU(3) gauge fields Xiao-Yong Jin

with the subscript 𝑐 denote the degree of freedom of the gauge Lie algebra, and per link (𝑥, 𝜇)
l2-norm as,
! 1/2
∑︁
2
Δ 𝑥, 𝜇 = 𝜕𝑥, 𝜇,𝑐 𝑆FT (𝑉; 𝛽) − 𝜕𝑥, 𝜇,𝑐 𝑆(𝑉; 𝛽T ) . (17)
𝑐

We choose the exponential of this l2-norm per link as the loss function for optimization:
∑︁

𝐿 LMEN = log exp Δ 𝑥, 𝜇 − log(4Vol). (18)
𝑥, 𝜇

For training the models, we generate configurations at 𝛽 = 0.7796, with a lattice size of 83 × 16,
using 4 HMC streams, and a trajectory length of 4, saving configurations every 16 trajectories, or 64
molecular dynamic time units, which is larger than the integrated autocorrelation lengths reported in
reference [31]. The computed topological charges show negligible autocorrelation. During training,
we use the target 𝛽T = 0.7099, corresponding to 𝑎/𝑟 0 ≃ 0.6, from an extrapolation formula in
Ref [30].
In a four-dimensional lattice, our code currently implement the transformation, using a stack of
smearing layers, each using the derivatives of 6 plaquette and 48 chair terms for the smearing kernel
per updating gauge link. The coefficients can be globally tuned numbers (6 + 48 = 54 parameters for
a single link) or the output of a constructed neural network as described in the previous section that
depends on nearby traced Wilson loops. In the following, we show two different models of gauge
field transformation: one with global coefficients (GC) and one with a Dense-Normalization-Dense
neural network (NN), each consisting of a stack of layers.
The order of layers in the GC model in Python’s list comprehension is:
transform=transform.TransformChain(
[StoutSmearSlice(coeff=CoefficientVariable(p0, chair=c0, rng=rng), dir=dir, is_odd=eo)
for _ in range(3) for dir in range(4) for _ in range(3) for eo in {False,True}]),

The layers update the gauge link in a single direction for the even and then odd subset 3 times.
These are repeated for 4 different directions, covering the whole lattice gauge links. Overall, the
whole lattice is updated 3 times. With 54 real parameters per layer, this transformation model uses
54 × 2 × 3 × 4 × 3 = 3888 parameters. The number of parameters is independent of the lattice size.
The order of layers in the NN model is:
transform=transform.TransformChain(
[StoutSmearSlice(coeff=CoefficientNets([Dense(units=8, activation=’swish’),
Normalization(),
Dense(units=54, activation=None)]), dir=dir, is_odd=eo)
for _ in range(4) for dir in range(4) for _ in range(1) for eo in {False,True}]

with the coefficients generated by a Dense network, Normalization, and another Dense network.
The units in the Dense network initialization denotes the dimension of the output vector. The last
Dense network does not use a non-linear activation function, as its 54 output numbers directly feed
into the input of the tan−1 function in equation (5).
We train our models using a single Nvidia A100 GPU with 40 GB memory. Training the GC
model uses 64 gauge configurations for 8 epochs, with each epoch going through the 64 configurations
in a random order. It takes 3 hours and uses a maximum of 30 GB GPU memory. The NN model,

6
Neural Network Gauge Field Transformation for 4D SU(3) gauge fields Xiao-Yong Jin

1e-03 1e-04
0.00095Δ𝐻1/3 0.0000763Δ𝐻1/3
HMC HMC
GC GC
NN NN

1e-04 1e-05

Δ𝐸(𝜏 = 7.28)
Δ𝐸(𝜏 = 4)

1e-05 1e-06

1e-06 1e-07
1e-07 1e-06 1e-05 1e-04 1e-03 1e-02 1e-07 1e-06 1e-05 1e-04 1e-03 1e-02

Δ𝐻 Δ𝐻

(a) 83 × 16 (b) 123 × 24

Figure 3: The difference in energy density versus the difference in Hamiltonian after a single Omelyan step.
The energy density is at Wilson flow time, 𝜏 = 4 for 83 × 16 with 𝛽 = 0.7796, and 𝜏 = 7.28 for 123 × 24
with 𝛽 = 0.8895. The step sizes for the Omelyan 2MN integrator are 𝛿𝑡 = 0.0025, 0.005, 0.01, and 0.02, for
symbols from left to right.

on the other hand, has more parameters and uses 256 configurations for training over 16 epochs.
That takes 11 hours and uses a maximum of 20 GB GPU memory.
During evaluation on a single Nvidia A100 GPU with 40 GB memory, we use gauge configura-
tions generated with HMC at 𝛽 = 0.7796 for 83 × 16 lattices, and 𝛽 = 0.8895 for 123 × 24 lattices,
using the DBW2 gauge action, from separate HMC streams that are different from those used for
training. With a lattice size of 83 × 16, a single force evaluation of the effective action, equation (2),
with the GC model takes 1.6 seconds and uses a maximum of 10 GB GPU memory. A single force
evaluation with the NN model takes 1.1 seconds and uses a maximum of 8 GB GPU memory. To
compare, using the same framework, a single force evaluation of the DBW2 lattice action without
transformations takes less than 0.005 seconds and uses a maximum of 128 MB GPU memory. The
evaluation of the same models on the 123 × 24 lattices requires more memory than the 40 GB
provided by the single Nvidia A100 we use, and thus runs on CPUs.
Since our machine learning framework (TensorFlow) has limited ability to distribute lattice
gauge configurations across GPUs, and our models have no lattice size dependencies, we focus on
how the machine-learned model scales with lattice size.
To study the effects of field transformation on the molecular dynamics in HMC, we start from
thermalized configurations and perform a single step of the Omelyan 2MN integrator [32]. We
measure the energy density using the lattice field tensor computed at 1-loop order with four plaquette
loops [33] at Wilson flow time, 𝜏 = 4, and examine how the integration step changes the energy
density and the Hamiltonian. Figure 3 shows the difference in energy density versus the difference
in Hamiltonian, before and after the integration step. The three kinds of symbols are from the
lattice action without transformation (labeled “HMC”), the lattice effective action with the field
transformation with global coefficients (labeled “GC”), and the lattice effective action with the field
transformation with a Dense-Normalization-Dense neural network (labeled “NN”). The same kind

7
Neural Network Gauge Field Transformation for 4D SU(3) gauge fields Xiao-Yong Jin

of symbols from left to right correspond to four different step sizes, 𝛿𝑡 = 0.0025, 0.005, 0.01, and
0.02. With a single integration step, Δ𝐻 scales with 𝛿𝑡 3 , and Δ𝐸 scales with 𝛿𝑡, thus Δ𝐸 ∝ Δ𝐻 1/3 .
We fit the line through the HMC data and show it in the figure.
The left panel in figure 3 shows the model performance on 83 × 16 at 𝛽 = 0.7796, while the
right shows performance on 123 × 24 at 𝛽 = 0.8895, which has 𝑎/𝑟 0 ≃ 0.265 and 𝑎 = 0.1326(13)
following Ref [30, 31], roughly keeping the same physical volume. The trained model achieves
smaller Δ𝐸 and Δ𝐻 with the same step size. For 83 × 16 lattices, the models manage to gain
larger Δ𝐸 at much smaller Δ𝐻, indicating a potential factor of a few better than HMC without
field transformations, keeping Δ𝐻 constant while inducing a larger change in Wilson-flowed energy
density. In addition, the same models trained on 83 × 16 work well with the 123 × 24 lattices,
maintaining a decent speedup with the larger lattice volume.

5. Conclusion

We implement and optimize gauge field transformation models for reduced effective gauge forces.
Applying the machine-optimized models as a change of variables in HMC reduces the Hamiltonian
violation, Δ𝐻, and increases the difference in Wilson-flowed energy, Δ𝐸, in our preliminary results.
The improvements remain when we directly apply the models to lattice configurations larger
than the lattice volume used for training, indicating the transferability of the learned transformations.
With further improvements in evaluation metrics, frameworks, and transformation models, this
approach remains a valuable research direction for enhancing sampling efficiency in lattice QCD
simulations.

Acknowledgments

We thank Peter Boyle, Norman Christ, Sam Foreman, Taku Izubuchi, Luchang Jin, Chulwoo
Jung, James Osborn, Akio Tomiya, and other ECP collaborators for insightful discussions and support.
This research was supported by the Exascale Computing Project (17-SC-20-SC), a collaborative
effort of the U.S. Department of Energy Office of Science and the National Nuclear Security
Administration. Part of this work was done on a pre-production supercomputer with early versions
of the Aurora software development kit. This research used resources of the Argonne Leadership
Computing Facility, a U.S. Department of Energy (DOE) Office of Science user facility at Argonne
National Laboratory and is based on research supported by the U.S. DOE Office of Science-Advanced
Scientific Computing Research Program, under Contract No. DE-AC02-06CH11357. We gratefully
acknowledge the computing resources provided and operated by the Joint Laboratory for System
Evaluation (JLSE) at Argonne National Laboratory.

References

[1] G. Cybenko, Approximation by superpositions of a sigmoidal function, Math. Control Signals

Syst. 2 (1989) 303.

[2] K. Hornik, M. Stinchcombe and H. White, Multilayer feedforward networks are universal
approximators, Neural Networks 2 (1989) 359.

8
Neural Network Gauge Field Transformation for 4D SU(3) gauge fields Xiao-Yong Jin

[3] K. Hornik, Approximation capabilities of multilayer feedforward networks, Neural Networks 4

(1991) 251.

[4] M. Leshno, V.Y. Lin, A. Pinkus and S. Schocken, Multilayer feedforward networks with a
nonpolynomial activation function can approximate any function, Neural Networks 6 (1993)
861.

[5] T. Cohen, M. Weiler, B. Kicanaoglu and M. Welling, Gauge equivariant convolutional

networks and the icosahedral CNN, in Proceedings of the 36th International Conference on
Machine Learning, K. Chaudhuri and R. Salakhutdinov, eds., vol. 97 of Proceedings of
Machine Learning Research, pp. 1321–1330, PMLR, 09–15 Jun, 2019,
https://proceedings.mlr.press/v97/cohen19d.html [1902.04615].

[6] M. Finzi, S. Stanton, P. Izmailov and A.G. Wilson, Generalizing convolutional neural
networks for equivariance to lie groups on arbitrary continuous data, in Proceedings of the
37th International Conference on Machine Learning, H.D. III and A. Singh, eds., vol. 119 of
Proceedings of Machine Learning Research, pp. 3165–3176, PMLR, 13–18 Jul, 2020,
https://proceedings.mlr.press/v119/finzi20a.html [2002.12880].

[7] D. Luo, G. Carleo, B.K. Clark and J. Stokes, Gauge Equivariant Neural Networks for Quantum
Lattice Gauge Theories, Phys. Rev. Lett. 127 (2021) 276402 [2012.05232].

[8] M. Favoni, A. Ipp, D.I. Müller and D. Schuh, Lattice Gauge Equivariant Convolutional Neural
Networks, Phys. Rev. Lett. 128 (2022) 032003 [2012.12901].

[9] Y. Nagai and A. Tomiya, Gauge covariant neural network for 4 dimensional non-abelian gauge
theory, 2103.11965.

[10] X.-Y. Jin, Neural Network Field Transformation and Its Application in HMC, PoS
LATTICE2021 (2022) 600 [2201.01862].

[11] J. Aronsson, D.I. Müller and D. Schuh, Geometrical aspects of lattice gauge equivariant
convolutional neural networks, 2303.11448.

[12] A. Tomiya and Y. Nagai, Equivariant transformer is all you need, PoS LATTICE2023 (2024)
001 [2310.13222].

[13] G. Kanwar, M.S. Albergo, D. Boyda, K. Cranmer, D.C. Hackett, S. Racanière et al.,
Equivariant flow-based sampling for lattice gauge theory, Phys. Rev. Lett. 125 (2020) 121601
[2003.06413].

[14] D. Boyda, G. Kanwar, S. Racanière, D.J. Rezende, M.S. Albergo, K. Cranmer et al., Sampling
using 𝑆𝑈 (𝑁) gauge equivariant flows, Phys. Rev. D 103 (2021) 074504 [2008.05456].

[15] S. Foreman, X.-Y. Jin and J.C. Osborn, Deep Learning Hamiltonian Monte Carlo, The Ninth
International Conference on Learning Representations, SimDL Workshop (2021)
[2105.03418].

9
Neural Network Gauge Field Transformation for 4D SU(3) gauge fields Xiao-Yong Jin

[16] D. Boyda et al., Applications of Machine Learning to Lattice Quantum Field Theory, in
Snowmass 2021, 2, 2022 [2202.05838].

[17] R. Abbott et al., Gauge-equivariant flow models for sampling in lattice field theories with
pseudofermions, Phys. Rev. D 106 (2022) 074506 [2207.08945].

[18] C. Lehner and T. Wettig, Gauge-equivariant neural networks as preconditioners in lattice

QCD, Phys. Rev. D 108 (2023) 034503 [2302.05419].

[19] C. Lehner and T. Wettig, Gauge-equivariant pooling layers for preconditioners in lattice QCD,
2304.10438.

[20] Y. Nagai and A. Tomiya, Self-learning Monte Carlo with equivariant Transformer,
2306.11527.

[21] S. Duane, A.D. Kennedy, B.J. Pendleton and D. Roweth, Hybrid Monte Carlo, Phys. Lett. B
195 (1987) 216.

[22] ALPHA collaboration, Critical slowing down and error analysis in lattice QCD simulations,
Nucl. Phys. B 845 (2011) 93 [1009.5228].

[23] G. Kanwar, Flow-based sampling for lattice field theories, in 40th International Symposium on
Lattice Field Theory, 1, 2024 [2401.01297].

[24] P.A. Boyle, Advances in algorithms for solvers and gauge generation, 2401.16620.

[25] M. Luscher, Trivializing maps, the Wilson flow and the HMC algorithm, Commun. Math. Phys.
293 (2010) 899 [0907.5491].

[26] C. Morningstar and M.J. Peardon, Analytic smearing of SU(3) link variables in lattice QCD,
Phys. Rev. D 69 (2004) 054501 [hep-lat/0311018].

[27] M. Abadi, A. Agarwal, P. Barham, E. Brevdo, Z. Chen, C. Citro et al., TensorFlow:

Large-scale machine learning on heterogeneous systems, 2015.

[28] P. Ramachandran, B. Zoph and Q.V. Le, Searching for activation functions, 2017.

[29] X.-Y. Jin, “Field transformation HMC with neural networks.” https://github.com/nftqcd/nthmc.

[30] S. Necco, Universality and scaling behavior of RG gauge actions, Nuclear Physics B 683
(2004) 137.

[31] G. McGlynn and R.D. Mawhinney, Diffusion of topological charge in lattice QCD simulations,
Phys. Rev. D 90 (2014) 074502 [1406.4551].

[32] I. Omelyan, I. Mryglod and R. Folk, Symplectic analytically integrable decomposition

algorithms: classification, derivation, and application to molecular dynamics, quantum and
celestial mechanics simulations, Computer Physics Communications 151 (2003) 272 .

[33] M. Lüscher, Properties and uses of the Wilson flow in lattice QCD, JHEP 08 (2010) 071
[1006.4518].

10