{"id":"9cce09d5-b6b3-44e0-b311-1d6cc0705ae3","arxiv_id":"2608.05437","paper_version":1,"verdict":"ACCEPT","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"For linear elastostatics, training a surrogate by minimizing discrete potential energy is exactly equivalent to supervised stiffness-norm regression, with identical gradients and no reference solutions.","lead":"Training a neural surrogate for linear elastic structures normally needs many reference solutions computed by the solver it is meant to replace. This note proves that minimizing the discrete potential energy of a prediction gives the same training signal as those reference solutions, with no labels at all.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No significant objection identified; the central exactness identity is algebraically sound and the main scope condition (same reduced DOF space, SPD K_ff) is explicitly stated.","rationale":"The reader's weakest assumption and mine coincide: the construction requires the surrogate to output vectors in the same reduced free-DOF space and requires K_ff SPD. However, I do not treat this as a correctness risk to the central claim, because the paper states the assumption in Section 2 ('Sufficient essential boundary conditions are assumed throughout') and reiterates in Section 9 that exactness is relative to the chosen discretisation. The quotient-space case is flagged but not developed; that is a limitation, not a refutation. The mathematical core is elementary and sound: the energy gap is exactly the squared stiffness-norm error, and the gradient is exactly the stiffness-weighted error. The surrounding results (Lemma 2, Corollary 1, the CG bound, Proposition 1) follow from standard arguments and are correctly scoped. The empirical result in Section 7 is explicitly 'not a claim of this note', and Section 8's dynamic claims are explicitly future work, so those passages cannot undermine the static central claim. The paper also provides executable checks and a pre-registration hash, which is independent support for the numeric claims. Therefore the ACCEPT verdict stands unchanged.","tokens_in":8723,"tokens_out":5422,"duration_ms":55939,"concrete_test":"Independently re-derive Lemma 1 from (1), then execute the published falsification suite at https://github.com/HymnOfLight/FE-JEPA and verify Table 1's measured values, in particular the Corollary 1 deviation 2.11e-15 and the Lemma 1 relative error ~1e-11. If either identity fails on any assembled SPD instance, the central claim is wrong.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is Lemma 1: for a fixed FE discretisation with SPD reduced stiffness K_ff, Π_h(u) - Π_h(U*) = (1/2)||u - U*||_K^2 and ∇Π_h(u) = K_ff(u - U*). The proof is two lines of exact algebra, and the identity is unconditional for every u in R^n. The only condition that could break the equivalence in practice is the requirement that the surrogate output live in the same reduced free-DOF space and that K_ff be SPD after eliminating Dirichlet constraints. Section 2 states this assumption explicitly and notes the quotient-space restatement for singular K_ff without developing it. That is a scope restriction, not an internal inconsistency: the paper repeatedly and correctly states that exactness is relative to the chosen discretisation and is not a claim about optimisation convergence, network expressivity, or generalisation. The empirical corroboration in Section 7 is explicitly labelled as context rather than a claim, and the elastodynamic extension in Section 8 is explicitly future work. I find no defect in the proof of the central claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","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.","tokens_in":8808,"tokens_out":10406,"duration_ms":90284,"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.","major_comments":[],"minor_comments":[{"comment":"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":"Section 6, Proposition 1"},{"comment":"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.","section":"Section 7"},{"comment":"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":"Table 1"},{"comment":"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.","section":"Section 6, scope counterexample"},{"comment":"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.","section":"Abstract and Section 7"}],"recommendation":"minor_revision","confidential_remarks":"The central exactness identity is algebraically sound and the paper is honest about its scope. The only mathematical blemish I found is the L_D=0 edge case in Proposition 1, which is easily fixed and does not affect the main claim. The contribution is modest but suitable for a note, and the executable verification and explicit counterexample are commendable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","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.\n\nThe 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.\n\nSoft 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.\n\nOverall: 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.","headline":"Correct but classical; the honest packaging and explicit scope make it worth a quick referee pass.","tokens_in":9442,"tokens_out":2676,"would_cite":true,"duration_ms":25492,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"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…","keywords":["finite element method","surrogate modelling","label-free training","minimum total potential energy","energy-norm conditioning","joint-embedding predictive architecture (JEPA)","executable falsification"],"falsifier":"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.","tokens_in":8399,"feed_emoji":"⚡","tokens_out":10927,"duration_ms":86763,"temperature":0.7,"pith_summary":"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.","feed_headline":"Discrete energy loss trains finite-element surrogates label-free","feed_subtitle":"Minimizing the assembled potential energy provably matches stiffness-norm regression toward the solver's own solution.","key_machinery":"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.","core_discovery":"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.","pith_inferences":["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."],"forward_implications":["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."],"supporting_citations":[{"why":"Supplies the classical energy-norm and SPD algebra on which the Lemma 1 identities rest.","marker":"[8]"},{"why":"Provides the Chebyshev bound for conjugate-gradient post-processing used in Section 5.","marker":"[9]"},{"why":"Introduces the variational energy-minimisation training route that the label-free objective builds on.","marker":"[1]"},{"why":"The variational operator learning baseline whose residual-based loss is compared with exact energy minimisation.","marker":"[5]"},{"why":"Defines the path toward autonomous machine intelligence that motivates the joint-embedding predictive architecture discussion.","marker":"[10]"},{"why":"Introduces the joint-embedding predictive architecture and the representational-collapse issue addressed by Proposition 1.","marker":"[11]"},{"why":"Supplies the conjugate-point theory used to quantify when the elastodynamic action is convex.","marker":"[12]"},{"why":"Newmark time integration whose implicit steps produce the effective SPD operator for step-by-step Lemma 1.","marker":"[13]"},{"why":"Generalized-$\\alpha$ integration, whose effective operator has the same form and inherits the result.","marker":"[14]"}],"fun_headline_variants":["Energy loss mimics stiffness-norm error without reference solutions","Exact label-free loss: energy equals half stiffness-norm error","Train FE surrogates directly on discrete energy, no solver runs","Energy-based loss provably matches stiffness-norm regression","No reference solutions needed: use discrete energy as loss"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"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.","fun_headline_variants_meta":{"raw":{"variants":["Energy loss mimics stiffness-norm error without reference solutions","Exact label-free loss: energy equals half stiffness-norm error","Train FE surrogates directly on discrete energy, no solver runs","Energy-based loss provably matches stiffness-norm regression","No reference solutions needed: use discrete energy as loss"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000623,"raw_usage":{"total_tokens":2934,"prompt_tokens":1045,"completion_tokens":1889,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":661,"completion_tokens_details":{"reasoning_tokens":1807}},"tokens_in":661,"tokens_out":1889,"duration_ms":13225,"temperature":1.0,"reasoning_tokens":1807,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T13:04:11.618606+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"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.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the classical energy-norm and SPD algebra on which the Lemma 1 identities rest."},{"cited_title":"Saad,Iterative Methods for Sparse Linear Systems, 2nd ed., SIAM, 2003","cited_arxiv_id":null,"evidence_quote":"Provides the Chebyshev bound for conjugate-gradient post-processing used in Section 5."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the variational energy-minimisation training route that the label-free objective builds on."},{"cited_title":"LeCun, A path towards autonomous machine intelligence, version 0.9.2, OpenReview preprint, 2022","cited_arxiv_id":null,"evidence_quote":"Defines the path toward autonomous machine intelligence that motivates the joint-embedding predictive architecture discussion."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the joint-embedding predictive architecture and the representational-collapse issue addressed by Proposition 1."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the conjugate-point theory used to quantify when the elastodynamic action is convex."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Newmark time integration whose implicit steps produce the effective SPD operator for step-by-step Lemma 1."},{"cited_title":"Chung and G","cited_arxiv_id":null,"evidence_quote":"Generalized-$\\alpha$ integration, whose effective operator has the same form and inherits the result."}],"review_version":1}