acceptodds
Under review as a conference paper at ICLR 2027

Convergence Is Not Accuracy: Transmission Conditions and Neural Training in FEM–PINN Coupling

Abstract

Hybrid solvers can compute one region of a domain with the finite-element method (FEM) and a neighboring region with a physics-informed neural network (PINN). The two solvers exchange boundary values and fluxes across their shared interface until the exchange settles. We show that when the PINN is inexact, the exchange can settle reliably on the wrong answer. For the classical Robin–Neumann exchange we prove that the Robin weight decides whether and how fast the exchange settles, while the answer it settles on is fixed by the PINN's interface response: for a fixed learned solver, every convergent Robin weight produces the same biased solution, with a closed-form error. Experiments on two-dimensional diffusion problems confirm this prediction to within . Couplings that pass the same stopping test differ by more than an order of magnitude in error, and enlarging the learned solver's feature space removes most of the gap. Classical solves stopped at a loose tolerance show the same premature stopping. Frequency probes trace the PINN's response error to both its features and its finite training budget, and the budget also changes how well the training loss tracks the coupled error. Hybrid couplings should therefore report the exchanged response and the physical error, not only the iteration progress.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
physics-informed neural networks, FEM-PINN coupling, domain decomposition, interface transmission conditions, Robin transmission, hybrid neural-numerical solvers, scientific machine learning, partial differential equations, convergence versus accuracy, spectral bias

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 4 reviews, above 27% of the others.
Reviewer_cfDFOpenReview ↗
Rating
2
Confidence
1
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This work couples FEM and PINN with a Robin-Neumann interface exchange, and argues that the Robin weight α\alpha control only whether and how fast the exchange converges, but not where it converges. The experiments check this on fixed random-feature solvers and on retrained PINNs. The paper also builds high frequency interface problems where all runs pass the stopping test with different errors, and track the coupled error by frequency.

Critical strengths
  • The writing is technically sound, and all the claims are well supported by formal proofs and experiments.
  • It compares against classical inexact solvers (CG), so it is fair to present inexactness in neural solvers.
  • The setup for experiments are measurable: with manufactured solutions and exact references.
Critical weaknesses
  • The novelty is limited. The main results follows from standard inexact fixed-point theory: no Robin weight can move the fixed point.
  • The "learning-specific" claim is overstated, since a coarse-mesh FEM solver can also settles on a wrong solution.
  • Although the paper can diagnoses failures in accuracy with manufactured solution or exact references, in real applications there is no such so there is no validated way to detect it.
  • For experiments, no retrained PINN runs on high-frequency problems, and the probes in section 4.3 are not linked to any coupled error.
Questions for authors
  • In section 4.3, the probes are Dirichlet problems, but in the coupling the PINN solvers a Neumann problem. So how can a network trained with Neumann condition solves a Dirichlet problem?
  • How does the learned solver’s error floor differ from a coarse-mesh FEM subsolver?
  • Do Dirichlet-probe responses predict the Neumann-mode sensitivity L of the trained PINN?
1 comment
AuthorsOpenReview ↗

Response to Reviewer cfDF

Thank you. Your coarse-FEM point was correct and improved the paper.

1. Novelty. We agree the theorem specializes classical inexact-solve analysis (F48x, point 1).

2. Coarse FEM (W2, Q2). The structure is identical: a coarse-mesh Neumann solver gives an α-free limit equal to Eq. (7) (≤7×10⁻⁸) and settles at up to 20% error. The difference is controllability: its trace bias falls about 4× when the mesh size is halved (24.3 → 5.8% at mode 8) and the global error returns to the reference on the reference mesh (20.4 → 7.3 → 3.56%, reference 3.57%), whereas a learned solver has no knob with a known rate: the restart policy alone changes Adam's error 3–4×, and training budget helps with no predictable rate. The claim is reworded accordingly.

3. Detection without a reference (W3). The held-out residual score needs no reference: on 214 settled runs it flags 130/132 inaccurate ones (33 false alarms in 82). We also show what it cannot see, a wrong interface flux (obDj, point 4).

4. Retrained PINNs at high frequency (W4). 120 runs: at mode 8 every run passes the update test at 100% trace error (77.7% global vs 3.57% FEM) for every α, flagged by the loss and the score; at mode 4 the runs never settle and their error grows with α; mode 1 and mode 4 with contrast 10 stay within 2.4× and 1.4× of FEM.

5. Probes (Q1, Q3). Each probe trains its own network per mode, so no network solves both problems. The Dirichlet responses predict the Neumann sensitivity only qualitatively (≈4× miss); the full measured L predicts the coupled limit exactly for fixed solvers.

Reviewer_2snCOpenReview ↗
Rating
2
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates the distinction between convergence and accuracy in FEM–PINN domain decomposition. For a Robin–Neumann exchange that reuses the imposed interface flux, the authors derive an error recurrence and a closed-form fixed-point bias for a fixed affine approximate subdomain response. Under the stated convergence conditions, the Robin parameter affects stability and convergence speed but not bias. Experiments on two-dimensional diffusion problems validate the prediction for a fixed response and show that couplings satisfying the same update-based stopping criterion can have substantially different errors. The experiments also include comparisons with inexact classical solvers and boundary-frequency probes to investigate the sources of interface-response error. Separate studies examine how the training budget and restart strategy affect the relationship between training loss and coupled-solution error. The study motivates assessing interface-response fidelity and physical error alongside iteration progress.

Critical strengths

A clear explanation of the gap between convergence and accuracy is provided. For the specified exchange with a fixed affine response, Theorem 1 explains why tuning the Robin parameter can improve convergence without removing the fixed-point bias.

Controlled experiments support the failure mode. Fixed-feature experiments validate the predicted limit, with higher-frequency cases showing substantial errors despite meeting the stopping criterion. Classical-solver controls further show that premature stopping is not unique to neural solvers.

Critical weaknesses

The limiting behavior of retrained PINNs needs further clarification. It remains unclear how closely the finite-budget retraining results reflect the limiting behavior described by the fixed-response analysis. Additional convergence evidence would strengthen the interpretation of Robin-parameter independence.

The relationship between the error range in Section 4.1 and the metrics in Figure 2(c) and Table 5 requires clarification. A consistent comparison would help assess the claimed insensitivity to the Robin parameter.

The connection between frequency probes and coupling errors requires further support. The separately trained Dirichlet probes provide useful diagnostics, but their relationship to the Neumann-to-trace response used in coupling needs further explanation.

Questions for authors

Q1

The updates in Table 5 remain nonzero after 25 exchanges. How can we tell whether the reported errors are close to stable limits that are largely independent of the Robin parameter? Longer runs or additional convergence checks would help clarify this. Comparisons using matched random seeds across Robin parameters could also help distinguish temporary training effects from persistent bias.

Q2

Please explain how the 0.61%–0.69% range in Section 4.1 relates to Figure 2(c) and Table 5. This range appears to refer to returned-trace errors, whereas the two-sided error medians range from 0.61% to 1.04%. Which metric supports the claim that the error changes little with the Robin parameter? Does considering both interface traces lead to the same conclusion?

Q3

How do the separately trained Dirichlet probes relate to the Neumann-to-trace response used in coupling? What evidence supports using these probes to explain the coupling errors observed in the experiments?

1 comment
AuthorsOpenReview ↗

Response to Reviewer 2snC

Thank you; your three questions each exposed something that needed fixing.

Q1. Are these limits? No, and the text now says "stochastic plateau". Over 100 exchanges and 10 seeds the updates stay at 0.39–0.68%. The per-run error plateaus at 0.65–0.67% with 250 steps and 0.34–0.43% with 1000; with 1000 steps it was still falling at exchange 25 (0.70%). With matched seeds, the plateau is α-insensitive at 250 steps and α=100 is worse by 0.10 pp [0.07, 0.14] at 1000 steps (two-sided metric: 0.12 pp [0.08, 0.17]). Separating the error of the mean trace from the fluctuation shows where that sits: the mean error is α-independent (0.12–0.15%), the Theorem 1 limit for a mean response error that is itself α-independent; the α effect is in the fluctuation, about 40% of it via the recursion's noise gain (0.92 → 1.08), the rest from noisier training at larger α.

Q2. Metric. You are right. The 0.61–0.69% is the returned-trace error; the two-sided error at exchange 25 is 0.61–1.04%. Both are now reported everywhere, with the updates alongside. The conclusion survives under both metrics in the form above: a 2× effect of the training budget, a ≤0.13 pp effect of α.

Q3. Dirichlet probes vs the Neumann coupling. They are not inverses for learned solvers: the product of the two learned responses falls to 0.00–0.07 at mode 8, and we deleted the sentence claiming otherwise. Neither probe predicts the per-mode coupled error quantitatively (typical miss ≈4× over 60 solver/problem pairs); the probes are now presented as qualitative diagnostics of the network class, and the mode-2 anomaly they revealed is explained (obDj, point 3).

Reviewer_obDjOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies a partitioned FEM–PINN coupling in which FEM solves a Robin problem on one subdomain and a PINN solves a Neumann problem on the other, exchanging interface data until the updates become small. The central claim is that convergence of the exchange is not accuracy. Theorem 1 gives an error recursion for the Robin data when the Neumann solver is replaced by a learned map N_k(q) = λ* + η_k − L_k(q − q*). For a fixed affine learned solver, the limit satisfies (I + L S_F)(λ̂ − λ*) = η and is independent of the Robin weight α, which affects only the rate. Experiments include 45 fixed random-feature couplings matching the predicted limit, controlled interface-mode problems where couplings passing the same stopping test differ greatly in error, a CG-based inexact-solver control, frequency probes of the transmitted flux, and a study of how training budget and warm/cold restarts change the loss–error relationship. The paper concludes with a reporting checklist for hybrid couplings.

Critical strengths

In hybrid numerical methods such as FEM–PINN coupling, small iterative updates are often used as an indicator of convergence. However, this paper demonstrates that convergence of the iterative process does not necessarily guarantee solution accuracy due to approximation errors in the learned solver. This finding highlights an important reliability issue in learning-based numerical methods and emphasizes the need to reconsider conventional convergence and stopping criteria.

Critical weaknesses

Theorem 1 is largely a specialization of existing inexact fixed-point analysis (Birken, 2015), and its theoretical novelty remains unclear. In particular, the independence of the limiting solution from the Robin parameter α\alpha in Theorem 1(b) follows directly from the surrogate interface balance equation, which does not involve α\alpha. More importantly, Theorem 1(b) assumes fixed LL and η\eta, which is valid for the random-feature least-squares solver but is not guaranteed for retrained PINNs. The boundedness assumptions of Theorem 1(a) are also not verified in the PINN experiments. Furthermore, Appendix C (Eq. 11) shows that flux mismatch introduces an explicit dependence on α\alpha, while Table 6 reports flux mismatches of 0.13–0.17. These limitations weaken both the theoretical contribution and its relevance to practical FEM–PINN coupling.

Also, the experiments are restricted to two-dimensional linear Poisson problems with circular interfaces and manufactured solutions. In contrast, prior FEM/SEM–PINN studies, such as Shukla et al. (2025), have considered nonlinear multiphysics problems, time-dependent flows, noisy observations, and real measurement data. The absence of comparable experiments makes it difficult to assess the generality of the reported failure mode. Moreover, the main "settled but wrong" examples involve severely inaccurate approximations, with global L2L^2 errors reaching 3.5–5.1. Appendix E (Fig. 7) further shows that a held-out residual diagnostic detects 30 of 31 inaccurate runs, suggesting that most failures can already be identified through conventional PDE residual checks. The paper also provides no concrete method to estimate or correct the learned-solver bias or establish a reliable stopping criterion. Consequently, the practical significance and broader applicability of the findings remain insufficiently demonstrated.

Questions for authors

-Sections 4.3–4.4 partially investigate the effects of optimization, but it remains unclear how changes in optimization strategies affect the assumptions and conclusions of Theorem 1. How do different optimizers (e.g., Adam vs. L-BFGS), training budgets, initialization strategies, and warm/cold restarts affect ηk\eta_k, LkL_k, and the final accuracy of the FEM–PINN coupling? In particular, under what conditions do the conclusions of Theorem 1 remain valid when the PINN is repeatedly retrained? Furthermore, how sensitive is the cap index defined in Section 4.3 to the threshold of 0.8? Do the conclusions remain consistent under longer or better-tuned optimization schedules beyond the single-seed experiments reported in Appendix I? Finally, is it possible to estimate η\eta during runtime, for example, through an additional probe solve at the current flux? Such an estimator could potentially extend the proposed analysis beyond a diagnostic framework toward a practical error estimation or correction method.

-Appendix E (Fig. 7) reports that the held-out PDE-residual diagnostic successfully identifies 30 out of 31 inaccurate runs. This suggests that most failures can already be detected using conventional PDE-residual checks. Does the proposed "settled but wrong" phenomenon therefore primarily demonstrate a limitation of update-based stopping criteria, or does it also persist when standard residual-based diagnostics are employed? Can the authors provide examples of failures that are not detected by conventional residual-based diagnostics, such as errors associated with η\eta that remain small under the PINN training loss but become apparent at held-out evaluation points? Alternatively, can the authors clarify what additional practical value the proposed analysis provides beyond existing residual-based diagnostic methods?

-The main numerical experiments are restricted to two-dimensional linear Poisson problems with circular interfaces and manufactured solutions. In contrast, previous FEM/SEM–PINN coupling studies, such as Shukla et al. (2025), have evaluated their methods on nonlinear multiphysics problems involving noisy and partially observed data. Can the authors provide evidence that the "settled but wrong" phenomenon and the independence of the limiting solution from the Robin parameter α\alpha also hold for nonlinear PDEs, non-circular interfaces, or time-dependent problems? When the interface exchange operator is no longer affine, which conclusions of Theorem 1 remain valid?

1 comment
AuthorsOpenReview ↗

Response to Reviewer obDj

Thank you for the most demanding review; several of your points changed the paper.

1. Novelty. Agreed: α-independence follows from the interface balance, and the revised text says so. See our reply to F48x, point 1, for what we claim instead. The nonlinear corollary answers your Q3: for any deterministic pair the fixed-point set is α-free; the closed-form bias is the affine part.

2. Theorem 1(a) and retrained PINNs. The hypotheses are checked where they hold (base problem: contraction factors 0.70–0.96) and stated where they fail (mode 4: up to 1.43; M_k changes too much between exchanges for any perturbation bound). The theorem is now a fixed-solver statement, verified with PINN-trained features; for retrained PINNs the mean coupling obeys it and α acts only on the variance (F48x, points 2–3).

3. Optimizers, restarts, threshold (Q1). 120 runs, 8 procedures: the error after 25 exchanges ranges from 0.44% (warm L-BFGS, α=10) to 3.38% (cold-restart Adam); α moves it ≤0.13 pp with cold restarts, up to 1.3 pp with warm-started L-BFGS. Cap ordering is unchanged at thresholds 0.7/0.8/0.9. The learned-MLP cap of 2 is one optimization failure at mode 2, not a cutoff: the Dirichlet probe of Table 1 has not yet fitted the boundary data there (mode 3 is transmitted), and 4× more steps raise the cap to 5. The Neumann probe fails at the same mode differently: it converges to the growing harmonic polynomial, which satisfies the PDE and the interface condition and violates only the exterior condition (loss 500× that of mode 1); 4× more steps remove it in 4 of 5 seeds. The representation limit is separate: features trained for 2500 steps cannot represent modes ≥4 (LS residual 0.98), and with 4× more steps mode 4 is transmitted as well. η at runtime is exact only with q*; the practical estimator is the score below.

4. Residual diagnostics (Q2). We adopt the score and quantify it: on 214 settled runs (random features, PINN-trained features, retrained PINNs) a score >1% flags 130/132 runs with error >1%, with 33 false alarms in 82. It is necessary: the mode-8 retrained PINNs pass the update test in 30/30 runs at 100% error. Its blind spot, as you anticipated: a solver with the right PDE and exterior condition but the wrong interface flux converges for every α, passes the update and two-sided checks, and keeps a score of 0.3–0.5% at 13% error; with an under-weighted Neumann condition the score decreases as the error grows. That is the error Theorem 1 locates at the interface; the flux jump sees it, relative to a nonzero baseline. This test is not current practice: the eight hybrid-coupling papers we checked stop on the interface update (Yin et al. 2022; Snyder et al. 2023, who add a check against the known solution; Sun et al. 2023, who already note that the outer iteration can converge with an erroneous flux), on the algebraic residual of the discretized system (HINTS, the classical case) or on a fixed budget (NeuroSEM; Nohra & Dufour).

5. Severity. The 3.5–5.1 cases (350–510% relative error) are the extreme rows; the same table has 69% and 15% at mode 8 (FEM 3.6%), and we add settled failures at 1.2–6% (general response, point 3).

6. Eq. (11) and Table 6. The §4.1 runs feed back the imposed flux, so the ζ/α term is absent from their recursion. The recorded mismatch is the network's own Neumann residual, set by cumulative training. Table 6 gives it in absolute terms (0.13–0.17 at exchange 25, against an RMS flux of 2.11), i.e. 6–8% after 25 000 Adam steps; it falls to 0.4–0.8% after 100 000 (exchange 100 of the new runs). The nodal mismatch is as large as the held-out one, so it is not an interpolation effect. It grows with α at equal budget because the imposed flux is noisier, and it does not alter the returned trace on which N_k is defined.

7. Scope. Heat, star-interface and semilinear results: F48x reply, point 4. NeuroSEM is now discussed; we do not address noisy or measured data.

Reviewer_F48xOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies hybrid solvers where one subdomain is solved with FEM and the neighboring one with a PINN, coupled through a Robin–Neumann exchange across their interface. The authors show that when the PINN is inexact, the exchange can converge to the wrong solution while still passing the usual stopping test. Their main result (Theorem 1) states that the Robin weight determines whether and how fast the exchange converges, but not the solution it converges to. For a fixed learned solver, every convergent Robin weight gives the same biased solution, with a closed-form error determined by the PINN’s interface response. They verify the theorem on 2D Poisson problems using fixed random-feature solvers and retrained PINNs. They show that couplings passing the same stopping test can differ in error by more than an order of magnitude. Frequency probes show that both the learned features and the training budget limit the PINN’s interface response. They also study how the training budget and restart policy affect the relationship between the training loss and the coupled error.

Critical strengths

Strength 1: Clarity of the paper.
The paper is well-written and easy to follow. The scalar example in the introduction (Appendix B) shows the main point in a few lines: the exchange settles in one step at the wrong value, and every other convergent Robin weight settles at the same wrong value. Figure 1 also explains the problem setup and the Robin–Neumann exchange clearly. This matters to my assessment because the main message of the paper is easy to understand.

Strength 2: Important practical message, supported by careful experiments.
The paper points out a pitfall in hybrid FEM–PINN couplings: passing the stopping test doesn’t mean the solution is accurate. The experiments support this well. The predicted limits match the measured ones closely, and the comparison with a classical conjugate gradient solver stopped at a loose tolerance separates the effects that are shared with classical inexact solvers from those specific to learned solvers. This is consequential because hybrid solvers are becoming more common, and the paper gives a clear reason to report the physical error and not only the convergence of the exchange.

Critical weaknesses

Weakness 1: Limited novelty and applicability of the theorem. In nonoverlapping domain decomposition, it is well known that the transmission condition affects the convergence rate but not the fixed point, since the fixed point is the solution of the coupled problem. Replacing the exact solver with an inexact one moves the fixed point, which follows from inexact fixed-point theory (Birken, 2015), as the authors acknowledge. It would help to state more clearly what the theorem adds beyond applying this framework to the FEM–PINN setting. The verification is also cleanest for the fixed-feature solver, which is a linear random-feature least-squares fit rather than a PINN. For the retrained PINNs, the learned map changes between exchanges, so the theorem doesn’t strictly apply, although the results are consistent with it. Extending the analysis to solvers that change between exchanges, or giving bounds that hold in that case, would make the connection to actual PINN couplings stronger. This is consequential because the theorem is the main contribution of the paper.

Weakness 2: Scope of the experiments and correcting the bias. The experiments use the 2D Poisson equation on a single geometry (a disc inside a square) with manufactured solutions. Testing other PDEs, such as time-dependent or nonlinear problems, or more complex geometries would show how general the conclusions are. This would be particularly relevant for fluid–structure interaction, which motivates the Robin–Neumann exchange. The paper would also benefit from a way to correct the bias once it’s detected. The held-out residual check is a promising start and could be developed further, for example by testing it on more problems. This is consequential because broader experiments and a remedy would make the paper’s findings more useful for practitioners.

Questions for authors
  1. The theorem applies exactly to the fixed-feature solver, but not to retrained PINNs, whose learned map changes between exchanges. Could the analysis be extended to this case, for example with bounds that hold when MkM_k and ηk\eta_k vary between exchanges?

  2. Have you tested other PDEs, such as time-dependent or nonlinear problems, or more complex geometries? Since the Robin–Neumann exchange is motivated by fluid–structure interaction, do you expect the same conclusions to hold there?

1 comment
AuthorsOpenReview ↗

Response to Reviewer F48x

Thank you for the careful reading and for naming the real gap: the theorem is clean only for the fixed solver.

1. Beyond inexact fixed-point theory. We agree it is a specialization and the text now says so. What we add: (i) the inexactness is persistent, not a tolerance that goes to zero, so the exchange converges to the exact fixed point of a different interface problem, given in closed form; (ii) the analysis must be done in Robin-data coordinates because cond(L) is 10⁸–10¹⁷ for learned solvers; (iii) which feedback designs keep α-independence (imposed flux yes, measured flux no, Eq. 11); (iv) a corollary: for any deterministic nonlinear pair the fixed-point set is α-free (verified on −Δu + cu³ to 5×10⁻⁸).

2. Not only random features. With PINN-trained features frozen and a least-squares head, 25/30 configurations converge for all three α; their limits coincide and match Eq. (7) to ≤7.5×10⁻⁸. At mode 4 they settle at 26–98% error (FEM 1.2%) for every α; at mode 8 they diverge for every α.

3. Varying M_k, η_k (your Q1). We measured η_k and L_k along 35 retrained trajectories with common-random-number probes. The exchange-to-exchange change of M_k (0.4–9) exceeds the contraction margin (≤0.09), so no perturbation bound is informative and we make no certification claim. What does hold is a stochastic reading: writing η_k = η̄ + ξ_k, the mean response error η̄ is α-independent (0.80–0.84% and 0.20–0.22% at the two budgets), so the mean of the retrained coupling has the α-free Theorem 1 limit with η̄. α enters only through the variance: the recursion's noise gain (0.92 at α=10, 1.08 at α=100) explains about 40% of the effect, noisier training at larger α the rest.

4. Other PDEs and FSI (your Q2). Heat equation: with converged exchanges per step the trajectory is α-free (1.5×10⁻⁷); with one exchange per step, the loosely coupled FSI practice, α enters through the splitting error (0.14–0.17% FEM–FEM, 0.6–1.4% with a PINN). Star interface and semilinear problem: α-free limits whenever the iteration converges; at the largest weight tested the star coupling diverges even FEM–FEM, the stability half of the theorem, which is what added-mass analyses address.

5. Remedy. The residual score (general response, point 4) plus feature enrichment: on the star and circle problems, 128→512 features brings the error to the FEM level.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General response: new results and revised PDF

We thank the reviewers. All four agree on the message, that an update-based stopping test does not certify accuracy, and all four ask the same three things: what the theorem adds beyond inexact fixed-point analysis, whether it says anything about retrained PINNs, and whether the failure is general and detectable. We ran 771 new configurations and three follow-up analyses. The main changes:

  1. Scope of the theorem, conceded and sharpened. Theorem 1 is stated for fixed solvers and we say plainly that α-independence follows from the interface balance. Verified to ≤1.5×10⁻⁷ for solvers with PINN-trained features, for coarse FEM, and on heat, star-interface and semilinear problems. A corollary extends the α-free fixed-point set to any deterministic nonlinear pair.

  2. Retrained PINNs: the theorem holds for the mean, α acts on the variance. Over 100 exchanges and 10 seeds, the error of the mean trace is α-independent (0.62–0.64% with 250 steps, 0.12–0.15% with 1000) because the network's mean response error is (0.80–0.84%, 0.20–0.22%); Theorem 1 then gives an α-free mean limit. The small α effect (per-run plateau 0.34→0.43% at 1000 steps) is entirely in the fluctuation: about 40% comes from the recursion's noise gain, the rest from noisier training at larger α. Training choices move the error 8× (0.44–3.38%); α moves it ≤0.13 pp in the paper's setting, up to 1.3 pp with warm-started L-BFGS.

  3. Settled but wrong with real PINNs, and milder cases. Retrained PINNs at interface mode 8 pass the update test in 30/30 runs at 100% trace error. Milder settled failures: PINN-trained features at mode 1 (3–6% vs 0.21% FEM), a semilinear case (3.5% vs 1.0%), and 12 retrained runs at 1.2–3.1%.

  4. The residual check becomes part of the method, with its blind spot shown. On 214 settled runs a held-out score above 1% flags 130/132 inaccurate runs (33 false alarms in 82 accurate ones). A solver with the wrong interface flux keeps a score of 0.3–0.5% at 13% error, so the score is a necessary check, not a certificate; the interface flux jump is its complement. None of the eight hybrid-coupling papers we checked uses such a test; they stop on the interface update, an algebraic residual (HINTS) or a fixed budget.

  5. Corrections. §4.1 reported the returned-trace metric only (two-sided: 0.61–1.04%); the retrained errors are plateaus, not limits; the "learning-specific" claim is reworded (coarse FEM has the same structure, but a refinement rate); the probe-inversion sentence is deleted; the learned-MLP cap of 2 is one optimization failure at mode 2, not a cutoff.

The revised PDF is uploaded. Changed text is coloured, the main text uses the extra page, Appendix A lists the changes, and Appendix P holds the new experiments with their tables.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.