Discrete energy as an exact label-free training objective for finite-element surrogates

Ruifeng Cao (The University of Manchester); Xidan Song (Wuhan University)

Discrete energy as an exact label-free training objective for finite-element surrogates

T0 review · 0 major / 5 minor · reviewed 2026-08-08 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read For linear elastostatics, the assembled discrete potential energy is an exact label-free training objective: minimizing it is supervised regression in the stiffness norm toward the solver's own solution, with identical gradients at every…

desk verdict Correct but classical; the honest packaging and explicit scope make it worth a quick referee pass. read the letter →

arxiv 2608.05437 v1 pith:5WXNBF2V submitted 2026-08-05 cs.CE cs.LGcs.NAmath.NA

Ruifeng Cao (The University of Manchester) , Xidan Song (Wuhan University) This is my paper · ORCID

classification cs.CEcs.LGcs.NAmath.NA

keywords finiteelementmethodsurrogatemodellinglabel-freetrainingminimumtotalpotentialenergyenergy-normconditioningjoint-embeddingpredictivearchitecture(JEPA)executablefalsification

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This note aims to break the circular cost of training finite-element surrogates, where every training label is a reference solution produced by the solver the surrogate should replace. For linear elastostatics, it proves that the assembled discrete potential energy $\Pi_h(u) = \frac{1}{2} u^\top K_{ff} u - F^\top u$ is an exact label-free objective: minimizing it is exactly supervised regression in the stiffness norm toward the solver's own solution, with identical gradients at every point. The proof is a short algebraic identity, evaluated using only the assembled stiffness operator and load vector, and the unknown energy of the solution is an irrelevant constant. A sympathetic reader should care because, if the claim holds, static finite-element surrogate training no longer needs a labelled corpus, and the same identity reorganizes evaluation, post-processing, and representation learning. The paper is explicit that exactness means exactness of the objective relative to the chosen discretisation, not a claim about optimization convergence, network expressivity, or generalization.

What carries the argument

The load-bearing object is the assembled discrete potential energy functional $\Pi_h(u) = \frac{1}{2} u^\top K_{ff} u - F^\top u$ on the reduced free-degree-of-freedom space of a fixed finite-element discretisation of linear elastostatics, where $K_{ff}$ is the symmetric positive definite stiffness operator and $F$ is the load vector. It works through the two identities of Lemma 1: the energy gap equals half the squared stiffness-norm error, and the gradient equals the stiffness-weighted error; equivalently, the gap equals the residual measured in the $K_{ff}^{-1}$-norm, with the inverse cancelling algebraically so the loss is evaluable by assembly alone, without a solve or a preconditioner. This functional is what converts training, evaluation, post-processing, and representation learning into exact statements on the reference discretisation, under the standing assumption that Dirichlet conditions remove rigid-body modes.

What would settle it

Compute, on any assembled linear elastostatic problem with symmetric positive definite $K_{ff}$ and load $F$, the residual $\Pi_h(u) - \Pi_h(U^*) - \frac{1}{2}\|u - U^*\|_K^2$ and $\|\nabla\Pi_h(u) - K_{ff}(u - U^*)\|$ for several random vectors $u$ and meshes with condition numbers up to $10^5$; if any instance deviates from zero beyond numerical tolerance (the paper reports $\lesssim10^{-11}$), Lemma 1 is false. Alternatively, train a surrogate on $\Pi_h$ alone on a validation split and test whether the stiffness-norm error decreases monotonically as the energy gap decreases.

Watch

Extended reading notes

Core claim

The central discovery is the two-line identity of Lemma 1: for every displacement vector $u$ in the reduced free-degree-of-freedom space, $\Pi_h(u) - \Pi_h(U^*) = \frac{1}{2}\|u - U^*\|_K^2$ and $\nabla \Pi_h(u) = K_{ff}(u - U^*)$. Because $U^* = K_{ff}^{-1} F$, the functional is assembled from mesh, material, and load data alone, so no reference solution is required. Consequently, the label-free objective $\Pi_h$ and supervised regression in the stiffness norm share the unique minimizer $U^*$ and have identical gradients at every point. The note presents this exactness as a property of the minimum principle of linear elastostatics on the reference discretisation, and uses it to derive a conditioning bound tying displacement error to the energy gap, a modewise contraction identity for gradient descent, a Chebyshev bound for conjugate-gradient polishing, and a conditional latent-separation bound for representation learning.

Load-bearing premise

The equivalence assumes the surrogate's output is a vector in the same reduced free-degree-of-freedom space of a fixed finite-element mesh, and that the stiffness operator $K_{ff}$ is symmetric positive definite; if the mesh changes or rigid-body modes are not eliminated, $\Pi_h$ is undefined or singular and the identities fail.

Editorial extensions

If this is right

Training a surrogate by minimizing the discrete energy, with no reference solutions, provably minimizes the same objective as supervised regression in the stiffness norm, so a corpus of unlabelled instances can be used at assembly cost.
The energy gap becomes the primary evaluation metric: Euclidean displacement error is controlled by the gap only through the condition number $\kappa$, and at the measured $\kappa\in[4.8\times10^3, 1.5\times10^5]$ a small gap can coexist with a large relative $\ell^2$ error.
Gradient descent on the energy contracts each stiffness eigenmode error exactly by $1-\eta\lambda_i$, which explains why smooth, low-frequency modes dominate the displacement error and why supervision on displacement alone can leave large energy-norm stress errors.
A surrogate prediction can be polished by conjugate-gradient iterations, and the energy gap after $k$ iterations obeys the Chebyshev bound $gap_k/gap_0 \le 4\rho^{2k}$, so post-processing cost and accuracy are certified in advance.
For a fixed geometry, accurate decoding of the energy forces separated latent codes, preventing representational collapse in joint-embedding predictive architecture pretraining; across different geometries, the descriptor-free separation statement is refuted by an explicit counterexample.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

Extending beyond the paper: the same algebraic cancellation that makes the loss evaluable by assembly should carry over to any quadratic minimum principle with an SPD operator, so the exactness is likely to hold for other linear elliptic problems, not only elastostatics.
Extending beyond the paper: the modewise contraction identity implies that the optimizer landscape is mesh-dependent, since $\kappa$ grows as the mesh refines; energy-trained surrogates will likely need preconditioned or second-order updates to avoid slowdown on smooth modes, a question the paper does not address.
Extending beyond the paper: the cross-geometry counterexample suggests a concrete architectural fix: condition the latent codes on geometry descriptors and test whether a transported-metric separation bound holds; this is a natural testable extension of Proposition 1.
Extending beyond the paper: the dynamics section leaves the assembly-level checks of the time-discrete repair as future work; verifying Lemma 1 on the effective operator $\tilde A = K_{ff} + (1/\beta\Delta t^2)M + (\gamma/\beta\Delta t)C$ on a simple wave problem would be a direct test of the proposed repair.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Punchline: don't let the arXiv date fool you; this is a correct, clean note, but the main identity is classical completing-the-square in the discrete energy. The real value is the honest scoping and the systematic consequences drawn for surrogate training. It deserves a serious referee, not a desk reject.

The paper does well: Lemma 1 states that minimizing Π_h is equivalent to supervised regression in the K-norm, with identical gradients. This is two lines of algebra, and they prove it correctly. They then draw consequences that are genuinely useful: the energy gap is the right primary metric because displacements can be large in ℓ2 while energy is small (Lemma 2 and Corollary 1 give the κ-controlled bound and the exact mode-wise contraction), and the CG polishing bound (5) is correctly stated and checked. The JEPA separation proposition is conditional, and they explicitly refute its cross-geometry extension with a measured counterexample. That kind of explicit scope discipline is rare and welcome. The executable falsification protocol, with thresholds fixed in advance and a pre-registration hash, is a real virtue; it makes the checks reproducible.

Soft spots: The novelty is modest—the identity is in any FE or iterative-solver textbook (refs [8,9] are properly cited). The empirical 'corroboration' in Section 7 is deliberately disclaimed as not a claim, and it shows: no error bars, no architecture details, just point estimates. That is fine as context but should not be read as evidence. The elastodynamic section is a summary of known results, not a proof, and the proposed repair via implicit time-stepping is plausible but not demonstrated with checks. The biggest practical caveat, which the paper states but does not discuss, is that the surrogate must output a vector in the same reduced DOF space; that limits the approach to fixed-mesh settings.

Overall: the central claim holds up. This is a useful reference for anyone training FE surrogates, and it clarifies the right evaluation metric. I would accept it for peer review with a request to make clear that the optimization difficulty for neural networks is not addressed.

Referee Report

0 major / 5 minor

Summary. The manuscript proves that for a fixed finite-element discretisation of linear elastostatics with a symmetric positive definite reduced stiffness operator K_ff, the discrete potential energy Pi_h(u)=1/2 u^T K_ff u - F^T u is an exact label-free training objective. Lemma 1 establishes Pi_h(u)-Pi_h(U*)=1/2 ||u-U*||_K^2 and grad Pi_h(u)=K_ff(u-U*), so that energy minimisation and supervised regression in the stiffness norm have the same unique minimiser and identical gradients with respect to u. The paper then derives a conditioning bound for the Euclidean displacement error (Lemma 2), a modewise contraction identity (Corollary 1), a conjugate-gradient polishing bound (Section 5), a conditional latent-separation proposition for JEPA pretraining on a shared operator (Proposition 1), and an argument that direct action minimisation is unsound for elastodynamics while an incremental time-discrete formulation restores exactness (Section 8). All numeric claims are implemented as executable falsification checks, with results reported in Table 1, and the exactness claim is explicitly limited to the reference discretisation rather than to optimisation convergence, network expressivity, or generalisation.

Significance. If the claims hold, the paper is a useful, clearly scoped contribution: it shows that for linear elastostatics on a fixed discretisation, the assembled energy can replace reference-solution labels at assembly cost, and it draws out practical consequences for evaluation metrics, CG post-processing, and representation learning. The core identity is elementary and follows from completing the square, but the paper's systematic treatment and its honest statement of scope are valuable. The executable falsification protocol and the explicit counterexample delimiting the JEPA separation claim are strengths that make the contribution more than a bare algebraic observation. The significance is moderate rather than groundbreaking, but it is appropriate for a note in this area.

minor comments (5)

[Section 6, Proposition 1] Equation (6) divides by L_D without assuming L_D>0. If L_D=0, the decoder is constant, and the accuracy hypothesis can hold only when ||U_i^*-U_j^*||_K <= 2 epsilon, in which case the right-hand side of (6) is undefined. Please either add the hypothesis L_D>0 or state the degenerate case separately.
[Section 7] The empirical corroboration reports displacement parity and a 5.1-fold energy-gap improvement but omits the network architecture, optimizer, hyperparameters, and mesh statistics for the experimental run. The code repository and hash are provided, but the text should either summarize these details or explicitly present the experiment as an illustrative demonstration rather than a full benchmark.
[Table 1] The row labelled 'Lemma 2 tightness' reports max ||e||_2 / sqrt(2 gap/\lambda_min) = 0.0224. This quantity measures how far the actual error is from the upper bound, not how tight the bound is; a value near 1 would indicate tightness. Please rename the quantity or clarify the intended interpretation.
[Section 6, scope counterexample] In the cross-geometry counterexample, the phrase 'on the same scale' is vague. Specify that the within-geometry comparison uses the same normalised Euclidean direction metric d, rather than the K-norm separation reported in Table 1 for the Proposition 1 premise.
[Abstract and Section 7] The statement that every claim with numeric content is checked as an executable falsification test is in tension with the paragraph stating that the empirical corroboration 'is not a claim of this note.' Please reword one of these passages, for example by saying 'every formal inequality claim' is checked.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity identified: the central exactness identity is an algebraic consequence of the definitions, with all scope conditions stated and all numeric claims independently checked.

full rationale

The derivation chain is self-contained and non-circular. Lemma 1 (Section 3) states the two identities Pi_h(u) - Pi_h(U*) = (1/2)||u - U*||_K^2 and grad Pi_h(u) = K_ff(u - U*), and the proof is immediate algebra: expand the squared K-norm and substitute K_ff U* = F. No fitted parameter, network output, or author-attributed prior result enters the identity; the only premises are the definition of the reduced FE system and the SPD property of K_ff, both stated explicitly in Section 2, including the caveat that rigid-body modes must be eliminated or the identities restated on a quotient space. Lemma 2, Corollary 1, the CG bound, and Proposition 1 are all derived from the same algebraic facts, with Proposition 1 treating epsilon and L_D as explicit hypotheses rather than inferred constants. The executable falsification checks of Section 7 compare measured quantities against predetermined inequalities and are described as pre-registered; the numerical parity result is explicitly labelled as context and not a claim of the note. The repository and pre-registration citations are independent support, not load-bearing argumentation. I find no step in which a prediction reduces by construction to its own input, no fitted input renamed as a prediction, and no uniqueness or ansatz imported through self-citation.

Assumptions & free parameters 0 free parameters · 7 assumptions · 0 invented entities

The central claim rests on standard linear algebra, not on fitted parameters. The main domain assumptions are the SPD property of the reduced stiffness operator and the surrogate operating in the same fixed free-DOF space. Proposition 1's Lipschitz and accuracy premises are explicitly conditional. The dynamics section relies on standard variational and time-integration results, but the omega_max = O(h^{-1}) asymptotic is not proven in the text.

assumptions (7)

domain assumption K_ff is symmetric positive definite after elimination of Dirichlet constraints
Section 2 assumes sufficient essential boundary conditions; if rigid-body modes exist, K_ff is singular and Lemma 1 must be restated on a quotient space.
standard math The discrete reference solution U* = K_ff^{-1} F is the unique minimizer of Pi_h
Follows from the SPD property and the minimum total potential energy principle; used in Lemma 1 and throughout.
ad hoc to paper The decoder D is L_D-Lipschitz with respect to the K-norm and achieves epsilon-accurate decoding
This is the premise of Proposition 1; the proposition is explicitly conditional on it, and the paper measures the separation premise but not the Lipschitz constant.
standard math The classical Chebyshev conjugate-gradient bound (5) holds with contraction factor rho
Bound (5) is cited from Saad [9] and used to certify the polishing step.
domain assumption The largest natural frequency scales as omega_max = O(h^{-1}) under mesh refinement
Section 8 uses this to conclude that the action's convexity horizon shrinks to zero in the continuum limit; stated without proof but standard for elliptic eigenvalue problems.
standard math The effective operator of Newmark-beta and generalized-alpha time stepping is a positive combination of K_ff, M, and C and is therefore SPD
Section 8 uses this to apply Lemma 1 step by step in the time-discrete repair.
standard math The second variation of the discrete action is positive semidefinite only for T <= pi / omega_max
Section 8 invokes the classical conjugate-point criterion from Gelfand and Fomin [12] to establish the elastodynamic obstruction.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Discrete energy as an exact label-free training objective for finite-element surrogates." pith.science (2026). https://pith.science/paper/5WXNBF2V

@misc{pith2026260805437,
  author       = {Pith},
  title        = {Pith review of: Discrete energy as an exact label-free training objective for finite-element surrogates},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5WXNBF2V}},
  note         = {Machine review of arXiv:2608.05437}
}

read the original abstract

Supervised training of finite-element (FE) surrogate models requires reference solutions, and each reference solution is obtained by solving the system that the surrogate is intended to replace. The assembled discrete potential energy provides a training signal that requires no reference solution. This note records, with proofs, the identities that make this signal exact for linear elastostatics: the difference between the energy of a prediction and the energy of the reference solution equals one half of the squared stiffness-norm error, and the gradient of the energy equals the stiffness-weighted error. Label-free discrete-energy minimisation and supervised regression in the stiffness norm therefore have the same unique minimiser and identical gradients at every point. Around this central result, the note states a conditioning lemma that bounds the displacement error by the energy gap, a modewise contraction identity that explains why the Euclidean displacement error is an unsuitable primary metric, the Chebyshev bound that governs conjugate-gradient post-processing of surrogate predictions, and a conditional latent-separation proposition for joint-embedding predictive architecture (JEPA) pretraining on a shared stiffness operator, with an explicit numerical counterexample that delimits its scope. Every claim with numeric content is implemented as an executable falsification check; the checks were executed twice, on synthetic test problems and on a probe set of 16 instances from the validation split of a pre-registered experimental run, and every inequality holds, with the measured tightness reported. A closing section explains why the construction does not extend to elastodynamics through direct minimisation of the action functional, and which time-discrete formulation restores exactness.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

19 extracted references · 18 canonical work pages

[1]

W. E and B. Yu, The Deep Ritz method: a deep learning-based numerical algorithm for solving variational problems,Commun. Math. Stat.6 (2018) 1–12

work page 2018
[2]

Samaniego, C

E. Samaniego, C. Anitescu, S. Goswami, V. M. Nguyen-Thanh, H. Guo, K. Hamdia, X. Zhuang and T. Rabczuk, An energy approach to the solution of partial differential equations in computational mechanics via machine learning: concepts, implementation and applications,Comput. Methods Appl. Mech. Engrg.362 (2020) 112790

work page 2020
[3]

Liao and P

Y. Liao and P. Ming, Deep Nitsche method: deep Ritz method with essential boundary conditions,Commun. Comput. Phys.29 (2021) 1365–1384

work page 2021
[4]

M. Liu, Z. Cai and K. Ramani, Deep Ritz method with adaptive quadrature for linear elasticity,Comput. Methods Appl. Mech. Engrg.(2023); arXiv:2305.15257

work page Pith review arXiv 2023
[5]

T. Xu, D. Liu, P. Hao and B. Wang, Variational operator learning: a unified paradigm marrying train- ing neural operators and solving partial differential equations,J. Mech. Phys. Solids(2024) 105714, doi:10.1016/j.jmps.2024.105714

arXiv 2024
[6]

Methods Appl

M.S.Eshaghi,C.Anitescu,M.Thombre,Y.Wang,X.ZhuangandT.Rabczuk,Variationalphysics-informed neural operator (VINO) for solving partial differential equations,Comput. Methods Appl. Mech. Engrg.437 (2025) 117785

work page 2025
[7]

An operator preconditioning perspective on training in physics-informed machine learning

T. De Ryck, F. Bonnet, S. Mishra and E. de Bézenac, An operator preconditioning perspective on training in physics-informed machine learning, in:Proc. 12th Int. Conf. on Learning Representations (ICLR), 2024; arXiv:2310.05801

work page Pith review arXiv 2024
[8]

P. G. Ciarlet,The Finite Element Method for Elliptic Problems, North-Holland, 1978

work page 1978

Show all 19 references

[9]

Saad,Iterative Methods for Sparse Linear Systems, 2nd ed., SIAM, 2003

Y. Saad,Iterative Methods for Sparse Linear Systems, 2nd ed., SIAM, 2003

2003
[10]

LeCun, A path towards autonomous machine intelligence, version 0.9.2, OpenReview preprint, 2022

Y. LeCun, A path towards autonomous machine intelligence, version 0.9.2, OpenReview preprint, 2022

2022
[11]

M.Assran,Q.Duval,I.Misra,P.Bojanowski,P.Vincent,M.Rabbat,Y.LeCunandN.Ballas,Self-supervised learningfromimageswithajoint-embeddingpredictivearchitecture,in:Proc.IEEE/CVFConf.onComputer Vision and Pattern Recognition (CVPR), 2023

2023
[12]

I. M. Gelfand and S. V. Fomin,Calculus of Variations, Prentice–Hall, 1963. 8

1963
[13]

N. M. Newmark, A method of computation for structural dynamics,J. Eng. Mech. Div. (ASCE)85 (1959) 67–94

1959
[14]

Chung and G

J. Chung and G. M. Hulbert, A time integration algorithm for structural dynamics with improved numerical dissipation: the generalized-αmethod,J. Appl. Mech.60 (1993) 371–375

1993
[15]

J. E. Marsden and M. West, Discrete mechanics and variational integrators,Acta Numerica10 (2001) 357–514

2001
[16]

Ortiz and L

M. Ortiz and L. Stainier, The variational formulation of viscoplastic constitutive updates,Comput. Methods Appl. Mech. Engrg.171 (1999) 419–444

1999
[17]

Brezis and I

H. Brezis and I. Ekeland, Un principe variationnel associé à certaines équations paraboliques,C. R. Acad. Sci. Paris282 (1976) 971–974; 1197–1198

1976
[18]

Nayroles, Deux théorèmes de minimum pour certains systèmes dissipatifs,C

B. Nayroles, Deux théorèmes de minimum pour certains systèmes dissipatifs,C. R. Acad. Sci. Paris282 (1976) A1035–A1038

1976
[19]

Serra and P

E. Serra and P. Tilli, Nonlinear wave equations as limits of convex minimization problems: proof of a conjecture by De Giorgi,Ann. of Math.175 (2012) 1551–1574. 9

2012

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.

[1] [1]

W. E and B. Yu, The Deep Ritz method: a deep learning-based numerical algorithm for solving variational problems,Commun. Math. Stat.6 (2018) 1–12

work page 2018

[2] [2]

Samaniego, C

E. Samaniego, C. Anitescu, S. Goswami, V. M. Nguyen-Thanh, H. Guo, K. Hamdia, X. Zhuang and T. Rabczuk, An energy approach to the solution of partial differential equations in computational mechanics via machine learning: concepts, implementation and applications,Comput. Methods Appl. Mech. Engrg.362 (2020) 112790

work page 2020

[3] [3]

Liao and P

Y. Liao and P. Ming, Deep Nitsche method: deep Ritz method with essential boundary conditions,Commun. Comput. Phys.29 (2021) 1365–1384

work page 2021

[4] [4]

M. Liu, Z. Cai and K. Ramani, Deep Ritz method with adaptive quadrature for linear elasticity,Comput. Methods Appl. Mech. Engrg.(2023); arXiv:2305.15257

work page Pith review arXiv 2023

[5] [5]

T. Xu, D. Liu, P. Hao and B. Wang, Variational operator learning: a unified paradigm marrying train- ing neural operators and solving partial differential equations,J. Mech. Phys. Solids(2024) 105714, doi:10.1016/j.jmps.2024.105714

arXiv 2024

[6] [6]

Methods Appl

M.S.Eshaghi,C.Anitescu,M.Thombre,Y.Wang,X.ZhuangandT.Rabczuk,Variationalphysics-informed neural operator (VINO) for solving partial differential equations,Comput. Methods Appl. Mech. Engrg.437 (2025) 117785

work page 2025

[7] [7]

An operator preconditioning perspective on training in physics-informed machine learning

T. De Ryck, F. Bonnet, S. Mishra and E. de Bézenac, An operator preconditioning perspective on training in physics-informed machine learning, in:Proc. 12th Int. Conf. on Learning Representations (ICLR), 2024; arXiv:2310.05801

work page Pith review arXiv 2024

[8] [8]

P. G. Ciarlet,The Finite Element Method for Elliptic Problems, North-Holland, 1978

work page 1978