arxiv: 2604.04321 · v1 · submitted 2026-04-06 · 🧮 math.DG · cs.LG

Recognition: no theorem link

Minimising Willmore Energy via Neural Flow

Edward Hirst , Henrique N. S\'a Earp , Tom\'as S. R. Silva

Authors on Pith no claims yet

Pith reviewed 2026-05-10 20:19 UTC · model grok-4.3

classification 🧮 math.DG cs.LG

keywords Willmore energyneural parameterizationmean curvatureminimal surfacesClifford torusPINN losssurface embeddingsgeometric flows

0 comments

The pith

Neural networks trained to minimize Willmore energy recover the round sphere for genus 0 and the Clifford torus for genus 1.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper presents a neural Willmore flow that uses neural architectures to represent embeddings of closed oriented surfaces in three-dimensional space. A physics-informed neural network loss is minimized directly on the Willmore energy, which is the integrated square of mean curvature. Training on genus-0 surfaces converges to the round sphere, while genus-1 surfaces converge to the Clifford torus, both known global minimizers. The same procedure is applied to genus-2 surfaces as a computational search strategy for an unresolved case in the Willmore problem.

Core claim

A neural parameterization of a surface embedding, evolved by minimizing a PINN-style loss on the Willmore energy, reproduces the round sphere when the surface has genus 0 and the Clifford torus when the surface has genus 1; the same procedure supplies a new numerical method for locating candidate minimal Willmore surfaces of genus 2.

What carries the argument

The neural Willmore flow: a map from a fixed 2-dimensional domain to R^3, realized by a neural network, whose parameters are updated to minimize a loss functional equal to the Willmore energy of the resulting embedding.

If this is right

The method identifies the known minimizers of the Willmore functional for genus 0 and genus 1 surfaces.
The same training procedure can be run on genus-2 surfaces to generate candidate minimal-energy embeddings.
The neural flow supplies a variational search technique for the open Willmore problem in higher genus.
Convergence of the loss to the known minimal values validates the use of PINN-style objectives for this geometric variational problem.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The approach could be tested on surfaces with prescribed topology or boundary conditions to locate other critical points of the Willmore functional.
Success on the Clifford torus suggests that similar neural losses might locate minimal-energy embeddings in other curvature-based problems such as total curvature minimization.
If the genus-2 output satisfies the Euler-Lagrange equation of the Willmore functional, it would constitute a new numerical candidate for the open problem.

Load-bearing premise

The neural network and the chosen loss function are expressive enough to locate the global minimizer of the Willmore energy rather than a local minimum or a discretization artifact.

What would settle it

Training the network on a genus-0 surface and obtaining a final embedding whose Willmore energy is strictly greater than 4 pi, or whose mean curvature is not constant, would show that the method does not reach the known minimizer.

Figures

Figures reproduced from arXiv: 2604.04321 by Edward Hirst, Henrique N. S\'a Earp, Tom\'as S. R. Silva.

**Figure 1.1.** Figure 1.1: The final trained embedded surfaces for each considered genus [PITH_FULL_IMAGE:figures/full_fig_p003_1_1.png] view at source ↗

**Figure 4.** Figure 4: then shows first a colouring of the fundamental domain of the sphere, which defines the ( [PITH_FULL_IMAGE:figures/full_fig_p012_4.png] view at source ↗

**Figure 4.1.** Figure 4.1: Training losses for the genus 0 setup. Willmore loss reaches the expected theoretical minimum for the round sphere; [PITH_FULL_IMAGE:figures/full_fig_p012_4_1.png] view at source ↗

**Figure 4.2.** Figure 4.2: Visualisation of the embedding evolution over training for the genus 0 setup. The fundamental domain is shown [PITH_FULL_IMAGE:figures/full_fig_p013_4_2.png] view at source ↗

**Figure 4.3.** Figure 4.3: Training losses for the genus 1 setup. Willmore loss reaches the expected theoretical minimum for the Clifford [PITH_FULL_IMAGE:figures/full_fig_p014_4_3.png] view at source ↗

**Figure 4.4.** Figure 4.4: Visualisation of the embedding evolution over training for the genus 1 setup. The fundamental domain is shown in [PITH_FULL_IMAGE:figures/full_fig_p014_4_4.png] view at source ↗

**Figure 4.5.** Figure 4.5: Training losses for the genus 2 setup. Willmore loss has a more erratic decline but drops to a stable region. [PITH_FULL_IMAGE:figures/full_fig_p015_4_5.png] view at source ↗

**Figure 4.6.** Figure 4.6: Visualisation of the embedding evolution over training for the genus 2 setup. The fundamentaxl domain is shown in [PITH_FULL_IMAGE:figures/full_fig_p016_4_6.png] view at source ↗

read the original abstract

The neural Willmore flow of a closed oriented $2$-surface in $\mathbb{R}^3$ is introduced as a natural evolution process to minimise the Willmore energy, which is the squared $L^2$-norm of mean curvature. Neural architectures are used to model maps from topological $2d$ domains to $3d$ Euclidean space, where the learning process minimises a PINN-style loss for the Willmore energy as a functional on the embedding. Training reproduces the expected round sphere for genus $0$ surfaces, and the Clifford torus for genus $1$ surfaces, respectively. Furthermore, the experiment in the genus $2$ case provides a novel approach to search for minimal Willmore surfaces in this open problem.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

Neural parameterization recovers the sphere and Clifford torus but the genus-2 candidate lacks the checks needed to confirm it is minimal.

read the letter

The paper applies a neural network to parameterize closed surfaces and minimizes the Willmore energy through a PINN-style loss on the mean curvature integral. It recovers the round sphere for genus 0 and the Clifford torus for genus 1, and it runs the same procedure on genus 2 as a computational probe for an open question. That is the core contribution: a fresh numerical search method for an old variational problem in surface geometry. The setup is straightforward and the known cases come out right, which shows the loss and architecture are at least functional for these topologies. The authors treat the genus-2 run as a new way to look for minimizers rather than a proof that one has been found. That distinction is honest and keeps the claim proportionate. The main weakness is the missing quantitative validation. There are no reported energy values with error estimates, no convergence diagnostics, no multiple random initializations, and no side-by-side comparison with existing discretizations such as finite-element or discrete differential geometry codes. Because the Willmore functional is non-convex, gradient-based training can stop at a local critical point whose energy is higher than the global minimum; nothing in the method rules that out for genus 2. The abstract alone does not supply enough detail to judge whether the learned surface is close to the theoretical lower bound or simply an artifact of the chosen network and quadrature. This work is aimed at people who already care about numerical methods for geometric variational problems or who follow the Willmore conjecture for higher genus. A reader looking for a new theorem or a fully validated computational result will find it thin, but someone exploring whether neural flows can serve as a practical search tool may get useful ideas. I would send it to peer review. The idea is clear, the classical cases work, and referees can ask for the missing diagnostics and comparisons without starting from zero.

Referee Report

3 major / 2 minor

Summary. The manuscript introduces a neural Willmore flow that parameterizes closed oriented surfaces in R^3 via neural maps from 2D topological domains and minimizes the Willmore energy ∫H² dA through a PINN-style loss. Training is reported to recover the round sphere (genus 0) and Clifford torus (genus 1), while the genus-2 experiment is presented as a new search method for minimizers of this open problem.

Significance. If the neural parameterization and optimization reliably attain the global infimum rather than local critical points, the approach would supply a flexible computational tool for the Willmore problem and related geometric variational problems, especially for higher-genus surfaces where analytic candidates are unavailable. The method's strength would lie in its mesh-free nature and potential extensibility, provided quantitative validation against known energies and convergence diagnostics are supplied.

major comments (3)

[Abstract / Experiments] Abstract and Experiments section: the claim that training 'reproduces the expected round sphere' and 'Clifford torus' is not supported by reported Willmore energy values, error bars, or direct comparison to the known analytic minima (4π for the sphere, 2π² for the Clifford torus). Without these diagnostics the recovery cannot be verified against the theoretical global minimum.
[Method / Experiments] Method and Experiments: the Willmore functional is non-convex on the space of immersions, yet the manuscript provides no evidence of multiple random initializations, convergence plots of the loss, or comparison against established discretizations (e.g., discrete Willmore flow or finite-element methods). This leaves open the possibility that reported shapes are local minima or discretization artifacts of the chosen network and quadrature.
[Experiments (genus 2)] Genus-2 experiment: the assertion that the procedure 'provides a novel approach to search for minimal Willmore surfaces' requires at least a quantitative comparison of the attained energy against previously published numerical upper bounds; absent such data the novelty claim rests only on the use of neural parameterization rather than on a demonstrably lower energy.

minor comments (2)

[Method] Notation for the neural map and the precise quadrature rule used to evaluate the PINN loss should be stated explicitly in the Method section to allow reproducibility.
[Method] The manuscript should clarify whether the neural parameterization enforces closedness and orientability by construction or via additional penalty terms.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive and detailed report. We address each major comment point by point below. The revisions will focus on adding the requested quantitative diagnostics and clarifying the scope of the genus-2 experiment without overstating the results.

read point-by-point responses

Referee: [Abstract / Experiments] Abstract and Experiments section: the claim that training 'reproduces the expected round sphere' and 'Clifford torus' is not supported by reported Willmore energy values, error bars, or direct comparison to the known analytic minima (4π for the sphere, 2π² for the Clifford torus). Without these diagnostics the recovery cannot be verified against the theoretical global minimum.

Authors: We agree that explicit numerical support is necessary for rigorous verification. The original manuscript presents the results primarily through visualizations of the recovered surfaces and the qualitative match to the known minimizers. In the revised version we will add a table reporting the final Willmore energies attained by the trained networks, together with standard deviations obtained from multiple independent training runs, and direct numerical comparisons against the analytic values 4π and 2π². revision: yes
Referee: [Method / Experiments] Method and Experiments: the Willmore functional is non-convex on the space of immersions, yet the manuscript provides no evidence of multiple random initializations, convergence plots of the loss, or comparison against established discretizations (e.g., discrete Willmore flow or finite-element methods). This leaves open the possibility that reported shapes are local minima or discretization artifacts of the chosen network and quadrature.

Authors: The non-convexity of the Willmore functional is a valid concern. Our experiments employed a deterministic initialization derived from the standard parametrization of the topological domain. We will augment the Experiments section with (i) loss-convergence plots for representative runs and (ii) results from at least five random initializations demonstrating that the same limiting shapes and energies are recovered. A short discussion comparing the neural energies to values reported in the discrete Willmore-flow literature will also be added; we note that a full side-by-side implementation of alternative discretizations lies beyond the scope of the present work but can be referenced for context. revision: partial
Referee: [Experiments (genus 2)] Genus-2 experiment: the assertion that the procedure 'provides a novel approach to search for minimal Willmore surfaces' requires at least a quantitative comparison of the attained energy against previously published numerical upper bounds; absent such data the novelty claim rests only on the use of neural parameterization rather than on a demonstrably lower energy.

Authors: The genus-2 experiment is presented as a proof-of-concept demonstration that the neural-flow framework can be applied to an open variational problem. We will revise the abstract and the relevant section to describe the result explicitly as a search procedure rather than a claim of a new global minimizer. In addition, we will locate the principal published numerical upper bounds for genus-2 Willmore surfaces and include a direct comparison of the energy obtained by our method against those bounds. revision: yes

Circularity Check

0 steps flagged

No significant circularity; direct minimization of classical functional

full rationale

The paper defines a neural parameterization of surfaces and minimizes a loss that is the direct discretization of the classical Willmore energy integrand (squared mean curvature) via standard PINN-style training. No equations or claims reduce by construction to fitted parameters renamed as predictions, self-definitional loops, or load-bearing self-citations. Reproduction of the round sphere and Clifford torus for genus 0/1 is presented as empirical validation against known analytic minimizers, while the genus-2 case is framed as an exploratory search without a uniqueness theorem or ansatz smuggled from prior author work. The derivation chain is therefore self-contained against the external mathematical definition of the Willmore functional.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The work rests on the classical definition of Willmore energy and on the standard PINN loss construction; no new free parameters or invented geometric objects are introduced in the abstract.

axioms (2)

standard math Willmore energy equals the L2 norm squared of mean curvature on a closed oriented surface in R3.
Standard definition in differential geometry, invoked in the abstract.
domain assumption A neural network can serve as a differentiable parameterization of an embedded surface whose energy can be minimized by gradient descent.
Core modeling choice of the method.

pith-pipeline@v0.9.0 · 5425 in / 1291 out tokens · 54266 ms · 2026-05-10T20:19:32.290347+00:00 · methodology

discussion (0)

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

A semicontinuous relaxation of Saito's criterion and freeness as angular minimization
math.AG 2026-04 conditional novelty 7.0

A new functional S vanishes precisely on free line arrangements and enables discovery of verified free examples for every admissible exponent pair with up to 20 lines.
PINNs in More General Geometry
math.DG 2026-04 unverdicted novelty 2.0

PINNs can address differential geometry problems by training neural networks to minimize functionals that encode geometric conditions, as shown through summaries of three related studies.

Reference graph

Works this paper leans on

44 extracted references · 13 canonical work pages · cited by 2 Pith papers · 2 internal anchors

[1]

Note on embedded surfaces,

T. J. Willmore, “Note on embedded surfaces,”An. S ¸tiint ¸. Univ. “Al. I. Cuza” Ia¸ si Sect ¸. I a Mat. (N.S.)11B(1965) 493–496

1965
[2]

Min-max theory and the willmore conjecture,

F. C. Marques and A. Neves, “Min-max theory and the willmore conjecture,”Annals of Mathematics 179no. 2, (2014) 683–782

2014
[3]

AInstein: Numerical Einstein Metrics via Machine Learning,

E. Hirst, T. S. Gherardini, and A. G. Stapleton, “AInstein: Numerical Einstein Metrics via Machine Learning,”arXiv:2502.13043 [hep-th]

work page arXiv
[4]

A Machine Learning Approach to the Nirenberg Problem,

G. Cort´ es, M. Esteban-Casadevall, Y. Feng, J. Henkel, E. Hirst, T. S. Gherardini, and A. G. Stapleton, “A Machine Learning Approach to the Nirenberg Problem,”arXiv:2602.12368 [cs.LG]

work page arXiv
[5]

Numerical Calabi–Yau metrics from holomorphic networks,

D. Michael, L. Subramanian, and Q. Yidi, “Numerical Calabi–Yau metrics from holomorphic networks,”Proceedings of Machine Learning Research145(2020) 223–252,arXiv:2012.04797 [hep-th]

work page arXiv 2020
[6]

Learning Size and Shape of Calabi-Yau Spaces,

M. Larfors, A. Lukas, F. Ruehle, and R. Schneider, “Learning Size and Shape of Calabi-Yau Spaces,” 11, 2021. 19

2021
[7]

arXiv , journal =:2211.12520 , journal =

M. Gerdes and S. Krippendorf, “CYJAX: A package for Calabi–Yau metrics with JAX,”Mach. Learn. Sci. Tech.4no. 2, (2023) 025031,arXiv:2211.12520 [hep-th]

work page arXiv 2023
[8]

Machine-learned Calabi–Yau metrics and curvature,

P. Berglund, G. Butbaia, T. H¨ uubsch, V. Jejjala, D. Mayorga Pe˜ na, C. Mishra, and J. Tan, “Machine-learned Calabi–Yau metrics and curvature,”Adv. Theor. Math. Phys.27no. 4, (2023) 1107–1158,arXiv:2211.09801 [hep-th]

work page arXiv 2023
[9]

Harmonic 1-forms on real loci of Calabi–Yau manifolds,

M. R. Douglas, D. Platt, and Y. Qi, “Harmonic 1-forms on real loci of Calabi–Yau manifolds,” May, 2024.https://arxiv.org/abs/2405.19402

work page arXiv 2024
[10]

https://arxiv.org/ abs/2602.12438

E. Heyes, E. Hirst, H. N. S. Earp, and T. S. R. Silva, “Neural and numerical methods for G2-structures on contact Calabi-Yau 7-manifolds,”arXiv:2602.12438 [math.DG]

work page arXiv
[11]

Computer-assisted proofs in pde: A survey,

J. G´ omez-Serrano, “Computer-assisted proofs in pde: A survey,”Notices of the AMS66no. 3, (2019) 298–310

2019
[12]

Asymptotic self-similar blow-up profile for three-dimensional axisymmetric Euler equations using neural networks,

Y. Wang, C.-Y. Lai, J. G´ omez-Serrano, and T. Buckmaster, “Asymptotic self-similar blow-up profile for three-dimensional axisymmetric Euler equations using neural networks,”Phys. Rev. Lett.130 (Jun, 2023) 244002.https://link.aps.org/doi/10.1103/PhysRevLett.130.244002

work page doi:10.1103/physrevlett.130.244002 2023
[13]

Non-uniqueness and symmetries for the nirenberg problem using computer assistance,

D. Platt, “Non-uniqueness and symmetries for the nirenberg problem using computer assistance,” 2026.https://arxiv.org/abs/2603.29544

work page arXiv 2026
[14]

Approximating high-dimensional minimal surfaces with physics-informed neural networks,

S. Zhou and X. Ye, “Approximating high-dimensional minimal surfaces with physics-informed neural networks,” 2023

2023
[15]

Physics-informed neural network solves minimal surfaces in curved spacetime,

K. Hashimoto, K. Kyo, M. Murata, G. Ogiwara, and N. Tanahashi, “Physics-informed neural network solves minimal surfaces in curved spacetime,”Mach. Learn. Sci. Tech.7no. 1, (2026) 015013, arXiv:2509.10866 [hep-th]

work page arXiv 2026
[16]

PyTorch: An Imperative Style, High-Performance Deep Learning Library

A. Paszke, S. Gross, F. Massa, A. Lerer, J. Bradbury, G. Chanan, T. Killeen, Z. Lin, N. Gimelshein, L. Antiga, A. Desmaison, A. K¨ opf, E. Yang, Z. DeVito, M. Raison, A. Tejani, S. Chilamkurthy, B. Steiner, L. Fang, J. Bai, and S. Chintala, “Pytorch: An imperative style, high-performance deep learning library,” 2019.https://arxiv.org/abs/1912.01703

work page internal anchor Pith review Pith/arXiv arXiv 2019
[17]

¨Uber konforme Geometrie I: Grundlagen der konformen Fl¨ achentheorie,

G. Thomsen, “ ¨Uber konforme Geometrie I: Grundlagen der konformen Fl¨ achentheorie,”Abh. Math. Semin. Univ. Hambg.3(1923) 31–56

1923
[18]

Blaschke,Vorlesungen ¨ uber Differentialgeometrie, III

W. Blaschke,Vorlesungen ¨ uber Differentialgeometrie, III. Springer, Berlin, 1929

1929
[19]

A global invariant of conformal mappings in space,

J. H. White, “A global invariant of conformal mappings in space,”Proc. Amer. Math. Soc.38no. 1, (1973) 162–164

1973
[20]

Metric flows with neural networks,

J. Halverson and F. Ruehle, “Metric flows with neural networks,”Mach. Learn. Sci. Tech.5no. 4, (2024) 045020,arXiv:2310.19870 [hep-th]

work page arXiv 2024
[21]

A new conformal invariant and its applications to the Willmore conjecture and the first eigenvalue of compact surfaces,

P. Li and S.-T. Yau, “A new conformal invariant and its applications to the Willmore conjecture and the first eigenvalue of compact surfaces,”Invent. Math.69no. 2, (1982) 269–291

1982
[22]

Existence of surfaces minimizing the willmore functional,

L. Simon, “Existence of surfaces minimizing the willmore functional,”Communications in Analysis and Geometry1no. 2, (1993) 281–326

1993
[23]

Existence of minimizing Willmore surfaces of prescribed genus,

M. Bauer and E. Kuwert, “Existence of minimizing Willmore surfaces of prescribed genus,”Int. Math. Res. Not.no. 10, (2003) 553–576

2003
[24]

Analysis aspects of Willmore surfaces,

T. Rivi` ere, “Analysis aspects of Willmore surfaces,”Invent. Math.174no. 1, (2008) 1–45

2008
[25]

Comparison surfaces for the Willmore problem,

R. Kusner, “Comparison surfaces for the Willmore problem,”Pacific J. Math.138no. 2, (1989) 317–345

1989
[26]

Complete minimal surfaces inS 3,

H. B. Lawson, “Complete minimal surfaces inS 3,”Ann. of Math. (2)92no. 3, (1970) 335–374. 20

1970
[27]

Minimizing the squared mean curvature integral for surfaces in space forms,

L. Hsu, R. Kusner, and J. Sullivan, “Minimizing the squared mean curvature integral for surfaces in space forms,”Experiment. Math.1no. 3, (1992) 191–207

1992
[28]

Goodfellow, Y

I. Goodfellow, Y. Bengio, and A. Courville,Deep Learning. MIT Press, 2016

2016
[29]

Learning representations by back-propagating errors,

D. E. Rumelhart, G. E. Hinton, and R. J. Williams, “Learning representations by back-propagating errors,”Nature323(1986) 533–536

1986
[30]

Adam: A Method for Stochastic Optimization

D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,”arXiv:1412.6980 [cs.LG]. https://arxiv.org/abs/1412.6980

work page internal anchor Pith review Pith/arXiv arXiv
[31]

Approximation by superpositions of a sigmoidal function,

G. Cybenko, “Approximation by superpositions of a sigmoidal function,”Math. Control Signals Systems2no. 4, (1989) 303–314

1989
[32]

Approximation capabilities of multilayer feedforward networks,

K. Hornik, “Approximation capabilities of multilayer feedforward networks,”Neural Netw.4no. 2, (1991) 251–257

1991
[33]

Universal approximation bounds for superpositions of a sigmoidal function,

A. R. Barron, “Universal approximation bounds for superpositions of a sigmoidal function,”IEEE Trans. Inform. Theory39no. 3, (1993) 930–945

1993
[34]

Neural network approximation,

R. DeVore, B. Hanin, and G. Petrova, “Neural network approximation,”Acta Numer.30(2021) 327–444

2021
[35]

Error bounds for approximations with deep ReLU networks,

D. Yarotsky, “Error bounds for approximations with deep ReLU networks,”Neural Netw.94(2017) 103–114

2017
[36]

Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations,

M. Raissi, P. Perdikaris, and G. E. Karniadakis, “Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations,” J. Comput. Phys.378(2019) 686–707

2019
[37]

Physics-informed machine learning,

G. E. Karniadakis, I. G. Kevrekidis, L. Lu, P. Perdikaris, S. Wang, and L. Yang, “Physics-informed machine learning,”Nat. Rev. Phys.3no. 6, (2021) 422–440

2021
[38]

Automatic differentiation in machine learning: a survey,

A. G. Baydin, B. A. Pearlmutter, A. A. Radul, and J. M. Siskind, “Automatic differentiation in machine learning: a survey,”J. Mach. Learn. Res.18no. 153, (2018) 1–43

2018
[39]

Finite elements on evolving surfaces,

G. Dziuk and C. M. Elliott, “Finite elements on evolving surfaces,”IMA J. Numer. Anal.27no. 2, (2007) 262–292

2007
[40]

Robust fairing via conformal curvature flow,

K. Crane, U. Pinkall, and P. Schr¨ oder, “Robust fairing via conformal curvature flow,”ACM Trans. Graph.32no. 4, (2013) 61:1–61:10

2013
[41]

Lawson surfaces

GeometrieWerkstatt, “Lawson surfaces.”http://geometriewerkstatt.com/LawsonSurfaces.html
[42]

C. M. Bishop,Pattern Recognition and Machine Learning. Information Science and Statistics. Springer, New York, NY, Aug., 2016

2016
[43]

Goodfellow, Y

I. Goodfellow, Y. Bengio, and A. Courville,Deep Learning. MIT Press, Cambridge, MA, 2016. http://www.deeplearningbook.org

2016
[44]

S. J. Prince,Understanding Deep Learning. The MIT Press, Cambridge, MA, 2023. http://udlbook.com. 21

2023