When is the combined load identifiable from a stress-intensity profile? A coupled forward-inverse study on SIFBench finite-element data

Filippo Grassia; Giansalvo Cirrincione

REVIEW 3 major objections 5 minor 11 references

For a known crack geometry, a combined tension–bending–bearing load is recoverable from a stress-intensity profile exactly when the three unit-load profiles form a non-degenerate triangle in function space; the paper proves this equivalence

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

For fixed crack geometry, combined tension–bending–bearing load is identifiable from the SIF profile iff the three elementary load profiles are affinely independent; recovery error is controlled by the minimum singular value of the profile matrix, not the condition number.

T0 review reviewed 2026-08-02 challenge →

load-bearing objection Solid identifiability theorem; the empirical 70/30 split and the calibration claim are both softer than the abstract suggests. the 3 major comments →

arxiv 2607.13074 v1 pith:WTEOQYLS submitted 2026-07-12 math.NA cond-mat.mtrl-scics.AIcs.LGcs.NA

When is the combined load identifiable from a stress-intensity profile? A coupled forward-inverse study on SIFBench finite-element data

Giansalvo Cirrincione , Filippo Grassia This is my paper

classification math.NA cond-mat.mtrl-scics.AIcs.LGcs.NA MSC 65J2065F35

keywords stress-intensity factorinverse problemsidentifiabilitycrack frontload mix recoverysimplex conditioningfinite-element benchmark

verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

The paper asks when the relative magnitudes of tension, bending, and bearing loads acting on a crack can be recovered from the stress-intensity-factor profile along the crack front. For a known geometry the forward map is exactly linear, so identifiability reduces to a single geometric question: whether the three elementary load profiles are linearly independent as functions along the front. The paper proves a closed-form characterization: a functional area that vanishes exactly in the unidentifiable regime, and shows that under noise the recovery error is controlled by the smallest singular value of a stability margin, not by the conditioning number alone. On a public finite-element benchmark of corner cracks, the typical geometry is well-posed (median simplex-restricted conditioning ≈ 20) while roughly 30% of geometries are genuinely ill-posed; a point estimate is therefore reliable on the majority and provably uninformative on the rest. The paper also contributes a forward surrogate and a calibrated set-valued inverse estimator, validated on controlled synthetic noise, and explicitly makes no forensic claim on real components.

Core claim

The central result, Theorem 1, is a four-way equivalence for a fixed geometry: the forward map is not injective on the affine hull of the load simplex; the functional area A(g) = √det G equals 0, where G is the Gram matrix of the difference profiles D1 = K_T − K_P and D2 = K_B − K_P; the three elementary load profiles are affinely dependent in L²(μ); and the smallest singular value of MB vanishes, where M is the profile matrix and B is an orthonormal basis of the sum-zero plane. Quantitatively, σ_min(MB)·σ_max(MB) = A(g)/√3 and the simplex-restricted conditioning κ_s(M) = √3 σ_max(MB)²/A(g), so κ_s diverges exactly as A(g) → 0. Proposition 3 proves that the recovery error under noise obeys ‖

What carries the argument

The profile matrix M(g) = [K_T | K_B | K_P] whose columns are the three unit-load profiles along the crack front, together with the functional area A(g) = √det G of the Gram matrix of the difference profiles. A(g) is twice the area of the triangle the three profiles span in function space; Theorem 1 ties it to the singular values of MB — σ_min·σ_max = A/√3 — so this single scalar simultaneously controls injectivity, conditioning, and noise-robustness. The same M(g) also serves as the differentiable forward map inside the inverse estimator, closing forward and inverse problems on one object.

Load-bearing premise

The empirical claims about the real dataset — the median κ_s ≈ 20, the ~30% ill-posed fraction, and the σ_min-controlled error correlations — are computed from the learned surrogate profiles, and the paper gives no sensitivity analysis establishing that κ_s and σ_min are stable to surrogate error.

What would settle it

Compute κ_s and σ_min directly from the exact finite-element profile matrices (not the learned surrogate) on the same held-out geometries and re-run the noise study; if the median κ_s, the ill-posed fraction, or the negative correlation between σ_min and error are not reproduced, the empirical transfer from surrogate to real profiles fails.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

A point estimate of (α,β,γ) is meaningful precisely when A(g) > 0; if A(g) = 0, no estimator can separate the loads and any point value is an artefact of the solver.
Under noise, recovery error scales as σ/σ_min(MB); the conditioning number κ_s can even anti-correlate with error across real geometries because σ_max varies, so uncertainty must be reported via σ_min.
On the public corner-crack data, the typical geometry is well-posed (median κ_s ≈ 20) but roughly 30% are ill-posed (κ_s > 50); the calibrated estimator widens its credible region along the degenerate direction in exactly those cases.
The shallow-crack collapse predicted by an analytical shape-function model is not observed systematically in the finite-element data; identifiability is governed by the full profile geometry, not relative depth alone.
With the true forward map, the inverse is exact at zero noise (error ~10⁻¹⁵ in every κ_s stratum), and the same σ_min-controlled ordering persists when the learned surrogate supplies M(g), showing the stability law survives surrogate approximation.

Where Pith is reading between the lines

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

The same A(g) criterion could serve as a cheap a priori screening tool: compute the profile triangle area once per geometry to flag load mixes that are structurally undecidable before any expensive inverse solve or experimental campaign.
Adding a second crack front or an independent measurement channel would likely restore identifiability in the ill-posed minority, because it increases the effective rank of M(g) and shrinks the degenerate direction — a natural candidate for the paper's proposed twin-crack extension.
Because the theory is model-free, it transfers to any linear-elastic superposition problem with an unknown coefficient vector — for example separating residual-stress contributions or mixed-mode SIF components — whenever the forward operator is exactly linear.
The empirical failure of the analytical shape-function prediction suggests that surrogate-based identifiability maps, computed from learned M(g), could be used to audit other analytic shape-function models, turning the released pipeline into a validation loop for the forward side.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

Solid identifiability theorem; the empirical 70/30 split and the calibration claim are both softer than the abstract suggests.

read the letter

Quick take: the theorem is correct and worth knowing; the applied packaging overreaches in two places that are fixable but real.

The genuinely new thing here is Theorem 1: an exact characterization of when the load mix is recoverable from the SIF profile, via a Gram determinant / functional area, plus the sigma_min vs kappa_s distinction in Proposition 3. The proof is clean, model-free, and the singular-value identity (sigma_min sigma_max = A/sqrt(3)) is a nice observation. The empirical demonstration that kappa_s alone anti-correlates with recovery error under noise, while sigma_min tracks it, is convincing on the true forward map (Test A'). I also give them credit for reporting the Newman-Raju failure on FEM data and explicitly bounding the scope to synthetic validation.

The soft spots are real, though. First, the calibration claim. The abstract and Section 6 say the estimator 'returns calibrated uncertainty' and that the claim is 'verified below'; Section 7.4 then says credible-region coverage is 'stated as a falsifiable prediction' rather than reported. That is a direct mismatch. The paper does not instrument the set-valued head, so the headline 'calibrated uncertainty' is not actually tested. It's fine to defer it, but the abstract should say so.

Second, the empirical identifiability split — median kappa_s ~20, ~30% ill-posed — is computed from the learned surrogate M-hat, not from the true FEM profiles, and no sensitivity analysis is given. Table 2 actually shows a warning sign: for the true map, 6000/20000 draws (30%) fall in the ill-posed stratum; for the learned map on the same geometries, only 3500/20000 (17.5%) do. That is a sizable discrepancy the paper does not discuss. Since kappa_s is built from differences between load profiles, surrogate bias on those differences could easily shift the 70/30 split. The authors have the true FEM profiles for the test set; they should have compared kappa_s(M-hat) with kappa_s(M_true) directly. This does not touch the theorem, but it does mean the dataset-level claim is not yet established.

Overall: this is a good, useful paper for anyone working on inverse fracture diagnostics or SIF surrogates, and it deserves a careful referee. The theory is solid, the exposition is honest about most limitations, and the sigma_min result is transferable. But the abstract overclaims calibration, and the empirical split needs a surrogate-bias check. I'd recommend peer review, with major revision.

Referee Report

3 major / 5 minor

Summary. The paper studies the inverse problem of recovering the tension/bending/bearing load mixture from a stress-intensity-factor profile along a crack front, using SIFBench finite-element data. For fixed geometry the forward map is exactly linear, and the paper's main theoretical contribution (Theorem 1) characterizes identifiability by the functional area A(g) of the triangle spanned by the three elementary load profiles in L^2: Φ is injective on the affine hull of the load simplex iff A(g)>0, equivalently σ_min(MB)>0, with quantitative identities σ_min(MB)σ_max(MB)=A(g)/√3 and κ_s=√3σ_max(MB)^2/A(g). Propositions 2 and 3 identify the degenerate direction and show that recovery error under noise is controlled by σ_min(MB), not by κ_s alone. A transformer-based crack-front operator is trained to supply M(g), and an inverse estimator with Dirichlet credible regions is described. The empirical sections report that on SIFBench corner-crack geometries the typical geometry is well-posed (median κ_s≈20) while ≈30% are ill-posed, that the noise behaviour follows the σ_min law, and that Test A (true forward map) recovers loads exactly at zero noise. The paper explicitly disclaims forensic use and states that credible-region coverage on the learned map is a falsifiable prediction rather than a reported result.

Significance. The theoretical core is sound and genuinely useful. Theorem 1 and Propositions 2–3 are derived cleanly from the linearity of the forward map; they give a closed-form, model-free criterion for identifiability and correctly identify σ_min(MB) as the intrinsic stability margin, explaining why κ_s alone can mislead when σ_max varies. The paper is also commendably honest: it reports that Corollary 4 (the Newman–Raju shallow-crack collapse) does not transfer to the FEM data, that the forward operator is behind RFR on the bending channel, and that no forensic validation is claimed. If the empirical claims about the SIFBench geometries — the 70/30 well/ill-posed split and the σ_min-controlled stability law — can be supported against the true FEM profiles rather than only the learned surrogate, and if the calibration claim is either verified or explicitly demoted, this would be a solid and citable contribution to inverse problems in fracture mechanics.

major comments (3)

[§7.3 and Table 2] The headline empirical identifiability statistics (median κ_s≈20.3, ≈30% ill-posed, Fig. 2) are computed from the learned surrogate M̂(g), not from the FEM profiles, as stated in §7.3. No comparison is reported between κ_s(M̂) and κ_s(M_FEM) on the same geometries, even though the true profiles are available (and are used in Test A). The discrepancy is visible in Table 2 / §7.4: Test A (true map) has 6000/20000 = 30% ill-posed draws, while Test B (learned M̂) has 3500/20000 = 17.5%; the text calls these 'matching' without discussing the factor-of-two difference. Because surrogate bias can systematically shift the κ_s distribution, the paper's central empirical claim that 'the typical geometry is well posed while a sizable minority is genuinely ill-posed' is not established for the actual SIFBench geometries. Please report κ_s(M_FEM) on the same test set and provide a sensitivity analysis
[§6 vs §7.4/§9] The central calibration claim is promised but not delivered. §6 states: 'the claim, verified below, is that the dispersion of the predicted posterior tracks the true non-identifiability measured by κ_s ... so that credible regions retain nominal coverage uniformly over the identifiability range.' Yet §7.4 says this claim 'is stated as a falsifiable prediction of the construction rather than reported here,' and §9 repeats that 'full credible-region coverage on the learned map is stated as a falsifiable diagnostic rather than reported.' This directly contradicts the abstract's 'calibrated uncertainty' and the 'verified below' in §6. Since calibrated uncertainty is a stated contribution (C3 and the inverse-estimator sections), this is a load-bearing omission. The authors should either add the coverage experiment on synthetic data (which they note is fully possible) or explicitly revise the
[§7.4 Test B / §7.3 Corollary 4 test] The claim that 'the stability law survives surrogate approximation' (corr(log σ_min(M̂), L1) ≈ −0.44 in Test B) is weakened by the same surrogate-bias issue: σ_min(M̂) is computed from the learned operator, and no true-map counterpart is given for the same geometries. Moreover, the negative finding about Corollary 4 — that κ_s shows no dependence on a/t on FEM data — is based on the surrogate-computed κ_s. If the surrogate smooths or distorts the depth dependence of the profiles, that negative conclusion could be an artifact. The true FEM profiles are available; a direct computation of κ_s and σ_min from M_FEM on the 1000 test geometries would settle both points and should be added.

minor comments (5)

[Abstract and C3] The phrase 'calibrated uncertainty' in the abstract and C3 overstates what is delivered; §7.4 explicitly defers coverage. Please align the wording with the actual content.
[§7.3] The first sentence says κ_s is computed 'over the test geometries,' but Figure 2 and the text then say 'over n_geom=4000 training geometries.' Please clarify which set is used.
[Table 2] The column 'B (M̂)' lacks a clear header for the counts; the n(A) column is explained only in the text. Also, the Test B stratum counts (12480/4020/3500) appear only in prose; adding them to the table would make the discrepancy visible and easier to discuss.
[Corollary 4] The statement 'the degenerate direction of Proposition 2 converges to (1,−1,0)' is stronger than the proof: the proof shows that (1,−1,0) is a near-null direction, not that the least singular vector converges to it. Please soften or provide a limit argument.
[Data and code availability] The text says the pipeline is 'released' (§3.3) but the final section says it 'will be deposited ... upon publication.' Please clarify the current availability; a preprint with a DOI or repository link would help reproducibility.

Circularity Check

0 steps flagged

No significant circularity: Theorem 1 is a self-contained linear-algebra characterization; the peripheral self-citation and self-referential synthetic validation are not load-bearing.

full rationale

The derivation chain is self-contained. Theorem 1 is proved by algebra from the definition of the forward map, the difference profiles D1, D2, and the Gram matrix; the equivalences (a)-(d) all reduce to whether D1, D2 are linearly independent, and the quantitative identities sigma_min*sigma_max = A/sqrt(3) and kappa_s = sqrt(3)*sigma_max^2/A follow from det([c1,c2]^T[c1,c2])=3 and the SVD of MB. No equation is assumed into the conclusion. Propositions 2 and 3 are consequences of the same singular-value calculation, not fitted results. Corollary 4 is explicitly conditional on the Newman-Raju model, and Section 7.3 reports that it fails on FEM data, which is an external falsifiable check rather than a circular confirmation. The empirical kappa_s distribution is computed from the learned Mhat rather than exact FEM profiles, and the synthetic noise study inverts profiles generated from the same linear model; the paper openly labels these as theory tests on synthetic data and disclaims forensic validation, so they are validation gaps, not circular reductions. The only author self-citation (Cirrincione 2026, incremental transformer) is peripheral and independently evidenced by the fixed-stack versus incremental comparison in Table 1. No load-bearing step reduces to its own input.

Axiom & Free-Parameter Ledger

1 free parameters · 6 axioms · 0 invented entities

No new physical entities are introduced. The 'functional area' A(g) and the simplex-restricted conditioning κ_s are defined mathematical scalars, not new forces, particles, or conserved quantities. The only hand-chosen numerical parameter affecting the reported empirical headline is the κ_s > 50 ill-posed threshold; all other inputs are standard LEFM assumptions or trained neural-network weights.

free parameters (1)

ill-posed stratum threshold κ_s > 50 = 50
Hand-chosen threshold used to define the 'ill-posed' stratum and to report the ≈30% ill-posed fraction. Not derived from data or theory; changing it changes the headline 70/30 split.

axioms (6)

domain assumption Linear-elastic superposition is exact for fixed geometry: K_mix = α K_T + β K_B + γ K_P (Eq. 1).
The entire linear forward-map and Theorem 1 rely on this. If crack-face contact, plasticity, or large deformation breaks superposition, the factorization fails.
domain assumption Combined loads are normalized to a convex mixture with α+β+γ=1.
The simplex constraint discards absolute load magnitude, so the theorem addresses relative load fractions only. If loads are not unit-normalized, a scale indeterminacy remains outside the stated identifiability result.
domain assumption Single-crack read-out of the twin-crack SIFBench scenario is valid; crack 2 enters only through its geometric parameters (a2/c2, a2/t).
The theory is developed for one crack. The paper assumes the second crack's influence on crack 1's field is fully captured by those covariates and does not invert the coupled two-crack inverse problem.
domain assumption The learned surrogate M̂(g) faithfully represents the true FEM profile matrix for computing κ_s and σ_min.
The empirical identifiability distribution and the σ_min correlations are computed from M̂, not the exact FEM profiles. No sensitivity analysis is provided to show the conclusions are robust to surrogate error.
standard math Standard linear algebra facts (SVD, Gram determinants, pseudoinverse norm bounds).
Used in Theorem 1 and Propositions 2–3; no issue.
domain assumption Newman–Raju shape functions in Corollary 4.
Corollary 4 is explicitly only under the Newman–Raju model and is tested and rejected on FEM data; it is not load-bearing for the central theorem.

reviewed 2026-08-02 · how reviews work

0 comments

Cite this review

Pith. "Pith review of When is the combined load identifiable from a stress-intensity profile? A coupled forward-inverse study on SIFBench finite-element data." pith.science (2026). https://pith.science/paper/WTEOQYLS

@misc{pith2026260713074,
  author       = {Pith},
  title        = {Pith review of: When is the combined load identifiable from a stress-intensity profile? A coupled forward-inverse study on SIFBench finite-element data},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WTEOQYLS}},
  note         = {Machine review of arXiv:2607.13074}
}

Share X Bluesky LinkedIn Reddit HN

read the original abstract

This work studies the inverse problem of recovering the relative magnitudes of the tension, bending, and bearing loads acting on a crack from its stress-intensity-factor profile along the crack front, using the public SIFBench finite-element data. The central claim is not forensic load recovery on field cases, but a rigorous characterization of when the combined load is identifiable at all, together with an estimator that returns calibrated uncertainty precisely in the regimes where it is not. For a known geometry the forward map from loads to profile is exactly linear, and identifiability reduces to a single geometric question: whether the three elementary load profiles are linearly independent as functions along the front. When they are nearly dependent, many different load combinations produce almost the same profile and the inverse problem is illposed; the analysis shows that the degree of ill-posedness is controlled by an intrinsic stability margin, not by the conditioning number alone. A single crack-front operator serves both as a structured forward surrogate and as the differentiable map required by a simplex-constrained, set-valued inverse estimator. On the SIFBench corner-crack scenario the empirical behaviour matches the theory: the typical geometry is well posed while a sizable minority is genuinely ill-posed, so a point estimate is reliable on the majority and provably uninformative on the rest. Validation is on controlled synthetic noise; no real fracture cases are used or claimed.

Figures

Figures reproduced from arXiv: 2607.13074 by Filippo Grassia, Giansalvo Cirrincione.

**Figure 1.** Figure 1: The three load pro les as points in function space. Left: non-degenerate triangle, the load is identi able. Right: near-collinear pro les, (𝑔) → 0, the load is not. Proof. (i). With Φ0 the linear part of Φ defined above, Φ is injective on the affine hull iff ker Φ0 ∩ {𝟏 ⊤𝑣 = 0} = {0}. For sum-zero 𝑣, 𝑣3 = −𝑣1 −𝑣2 and Φ0 (𝑣) = 𝑣1𝐷1 +𝑣2𝐷2 ; hence injectivity holds iff 𝐷1 , 𝐷2 are linearly independent, i.e. … view at source ↗

**Figure 2.** Figure 2: Empirical identi ability on the nite-element data (𝑛geom = 4000): distribution of log10 𝜅𝑠 (left) and log10 𝜅𝑠 versus 𝑎∕𝑡 (right). factor on this scenario exceeds 108 , so raw-magnitude regression is ill-conditioned. The operator additionally provides exact symmetry and exact positivity by construction, which no tabular regressor does, and yields the full 𝑀(𝑔) in one evaluation — the property the inverse s… view at source ↗

**Figure 3.** Figure 3: Test A′ : inverse error against log10 𝜅𝑠 (left) and against log10 𝜎min(𝑀𝐵) (right). 7.4. Inverse: accuracy and calibration The inverse is evaluated under a deliberate separation of concerns, so that a theory test is not confounded with surrogate quality. Test A solves the inverse from the true forward map (the FEM profiles themselves as 𝑀): noiseless, this must reconstruct (𝛼, 𝛽, 𝛾) exactly by Theorem 1, a… view at source ↗

**Figure 4.** Figure 4: Mean inverse 𝐿1 vs. noise 𝜎, one line per 𝜅𝑠 stratum. Extending to real cases would require components with independently known loading histories and measured front profiles; the calibrated estimator is designed so that, on such data, it would report not just a load estimate but the geometry-intrinsic certificate 𝜅𝑠 saying whether that estimate is constrained by the measurement. 9. Conclusion The study rea… view at source ↗

discussion (0)

Reference graph

Works this paper leans on

11 extracted references · 3 canonical work pages

[6]

Journal of Machine Learning Research 24, 1–97

Neural operator: Learning maps between function spaces. Journal of Machine Learning Research 24, 1–97. doi:10.48550/arXiv.2108.08481. Li, Z., Kovachki, N., Azizzadenesheli, K., Liu, B., Bhattacharya, K., Stuart, A., Anandkumar, A.,

Pith/arXiv arXiv doi:10.48550/arxiv.2108.08481
[1957]

Journal of Applied Mechanics 24, 361–364

Analysis of stresses and strains near the end of a crack traversing a plate. Journal of Applied Mechanics 24, 361–364. doi:10.1115/1.4011547. Kaipio, J., Somersalo, E.,

doi:10.1115/1.4011547
[1981]

Engineering Fracture Mechanics 15, 185–192

An empirical stress-intensity factor equation for the surface crack. Engineering Fracture Mechanics 15, 185–192. doi:10.1016/0013-7944(81)90116-8. Newman, J.C., Raju, I.S.,

work page doi:10.1016/0013-7944(81)90116-8
[1983]

ASTM International

Stress-intensity factor equations for cracks in three-dimensional finite bodies, in: Fracture Mechanics: Fourteenth Symposium—Volume I: Theory and Analysis. ASTM International. doi:10.1520/STP37074S. Stuart, A.M.,

work page doi:10.1520/stp37074s
[1991]

Engineering Fracture Mechanics 40, 1135–1146

Universal features of weight functions for cracks in mode I. Engineering Fracture Mechanics 40, 1135–1146. doi:10.1016/0013-7944(91)90177-3. Hadamard, J.,

work page doi:10.1016/0013-7944(91)90177-3
[2005]

Inverse Problem Theory and Methods for Model Parameter Estimation. SIAM. Vaswani,A.,Shazeer,N.,Parmar,N.,Uszkoreit,J.,Jones,L.,Gomez,A.N.,Kaiser,Ł.,Polosukhin,I.,2017. Attentionisallyouneed,in:Advances in Neural Information Processing Systems (NeurIPS). Vovk, V., Gammerman, A., Shafer, G.,

2017
[2010]

Acta Numerica 19, 451–559

Inverse problems: a Bayesian perspective. Acta Numerica 19, 451–559. doi:10.1017/S0962492910000061. Tancik, M., Srinivasan, P.P., Mildenhall, B., Fridovich-Keil, S., Raghavan, N., Singhal, U., Ramamoorthi, R., Barron, J.T., Ng, R.,

doi:10.1017/s0962492910000061
[2021]

Nature Machine Intelligence 3, 218–229

Learning nonlinear operators via DeepONet based on the universal approximation theorem of operators. Nature Machine Intelligence 3, 218–229. doi:10.1038/s42256-021-00302-5. Newman, J.C., Raju, I.S.,

doi:10.1038/s42256-021-00302-5
[2023]

Foundations and Trends in Machine Learning 16, 494–591

Conformal prediction: A gentle introduction. Foundations and Trends in Machine Learning 16, 494–591. doi:10.1561/2200000101. Cirrincione, G.,

doi:10.1561/2200000101
[2025]

arXiv preprint arXiv:2506.01173 ArXiv:2506.01173 [cs.DB]

SIFBench: An extensive benchmark for fatigue analysis. arXiv preprint arXiv:2506.01173 ArXiv:2506.01173 [cs.DB]. Glinka, G., Shen, G.,

Pith/arXiv arXiv
[2026]

arXiv preprint arXiv:2604.10703 doi:10.48550/ arXiv.2604.10703

INCRT: An incremental transformer that determines its own architecture. arXiv preprint arXiv:2604.10703 doi:10.48550/ arXiv.2604.10703. arXiv:2604.10703 [cs.LG]. Engl, H.W., Hanke, M., Neubauer, A.,

Pith/arXiv arXiv doi:10.48550/arxiv.2604.10703

This paper was first reviewed by deepseek-v4-flash on August 2, 2026.

[1] [6]

Journal of Machine Learning Research 24, 1–97

Neural operator: Learning maps between function spaces. Journal of Machine Learning Research 24, 1–97. doi:10.48550/arXiv.2108.08481. Li, Z., Kovachki, N., Azizzadenesheli, K., Liu, B., Bhattacharya, K., Stuart, A., Anandkumar, A.,

Pith/arXiv arXiv doi:10.48550/arxiv.2108.08481

[2] [1957]

Journal of Applied Mechanics 24, 361–364

Analysis of stresses and strains near the end of a crack traversing a plate. Journal of Applied Mechanics 24, 361–364. doi:10.1115/1.4011547. Kaipio, J., Somersalo, E.,

doi:10.1115/1.4011547

[3] [1981]

Engineering Fracture Mechanics 15, 185–192

An empirical stress-intensity factor equation for the surface crack. Engineering Fracture Mechanics 15, 185–192. doi:10.1016/0013-7944(81)90116-8. Newman, J.C., Raju, I.S.,

work page doi:10.1016/0013-7944(81)90116-8

[4] [1983]

ASTM International

Stress-intensity factor equations for cracks in three-dimensional finite bodies, in: Fracture Mechanics: Fourteenth Symposium—Volume I: Theory and Analysis. ASTM International. doi:10.1520/STP37074S. Stuart, A.M.,

work page doi:10.1520/stp37074s

[5] [1991]

Engineering Fracture Mechanics 40, 1135–1146

Universal features of weight functions for cracks in mode I. Engineering Fracture Mechanics 40, 1135–1146. doi:10.1016/0013-7944(91)90177-3. Hadamard, J.,

work page doi:10.1016/0013-7944(91)90177-3

[6] [2005]

Inverse Problem Theory and Methods for Model Parameter Estimation. SIAM. Vaswani,A.,Shazeer,N.,Parmar,N.,Uszkoreit,J.,Jones,L.,Gomez,A.N.,Kaiser,Ł.,Polosukhin,I.,2017. Attentionisallyouneed,in:Advances in Neural Information Processing Systems (NeurIPS). Vovk, V., Gammerman, A., Shafer, G.,

2017

[7] [2010]

Acta Numerica 19, 451–559

Inverse problems: a Bayesian perspective. Acta Numerica 19, 451–559. doi:10.1017/S0962492910000061. Tancik, M., Srinivasan, P.P., Mildenhall, B., Fridovich-Keil, S., Raghavan, N., Singhal, U., Ramamoorthi, R., Barron, J.T., Ng, R.,

doi:10.1017/s0962492910000061

[8] [2021]

Nature Machine Intelligence 3, 218–229

Learning nonlinear operators via DeepONet based on the universal approximation theorem of operators. Nature Machine Intelligence 3, 218–229. doi:10.1038/s42256-021-00302-5. Newman, J.C., Raju, I.S.,

doi:10.1038/s42256-021-00302-5

[9] [2023]

Foundations and Trends in Machine Learning 16, 494–591

Conformal prediction: A gentle introduction. Foundations and Trends in Machine Learning 16, 494–591. doi:10.1561/2200000101. Cirrincione, G.,

doi:10.1561/2200000101

[10] [2025]

arXiv preprint arXiv:2506.01173 ArXiv:2506.01173 [cs.DB]

SIFBench: An extensive benchmark for fatigue analysis. arXiv preprint arXiv:2506.01173 ArXiv:2506.01173 [cs.DB]. Glinka, G., Shen, G.,

Pith/arXiv arXiv

[11] [2026]

arXiv preprint arXiv:2604.10703 doi:10.48550/ arXiv.2604.10703

INCRT: An incremental transformer that determines its own architecture. arXiv preprint arXiv:2604.10703 doi:10.48550/ arXiv.2604.10703. arXiv:2604.10703 [cs.LG]. Engl, H.W., Hanke, M., Neubauer, A.,

Pith/arXiv arXiv doi:10.48550/arxiv.2604.10703

When is the combined load identifiable from a stress-intensity profile? A coupled forward-inverse study on SIFBench finite-element data

The reading

Core claim §

What carries the argument

Load-bearing premise §

What would settle it §

If this is right

Where Pith is reading between the lines

discussion (0)

Core claim

Load-bearing premise

What would settle it