acceptodds
Under review as a conference paper at ICLR 2027

Certifying Physics-Informed Surrogates for PDE-Constrained Optimization

Abstract

Physics-informed neural surrogates can reduce the cost of PDE-constrained optimization, but an optimizer itself can drive iterates into regions where the surrogate is least reliable. We introduce SA-CPITR, a solver-assisted certified trust-region method that treats a learned PDE surrogate as an inexact optimization oracle whose errors are bounded online. A posteriori majorants convert PDE residual information into computable error bounds for the state and sensitivities, which we assemble into certificates for the reduced objective and its gradient. Since the majorants hold for any conforming candidate, the same formulas certify both the neural surrogate and the high-fidelity solver outputs. These certificates control trust-region acceptance and termination, guaranteeing objective descent for every accepted step without evaluating the true objective. When a certificate becomes insufficient, the method spends state or gradient high-fidelity solves to generate a reusable anchor that tightens subsequent certificates. If the target tolerance cannot be certified within the anchor event budget, the algorithm returns its best certified bound rather than an uncertified solution. We prove certificate validity, guaranteed descent, and convergence to first-order criticality under explicit assumptions. On five PDE control problems, from linear elliptic to nonlinear and hyperbolic classes, SA-CPITR certifies terminal criticality on all five with a total of only three anchor events, while the corresponding solver-free variant certifies two. A sixth problem extends the method to state constraints, with certified feasibility and descent. The experiments use a fixed quadrature grid to evaluate integrals inside the optimization loop, and then independently re-verify certificates at every returned point by exact integration or continuum-valid enclosures.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
physics-informed neural surrogate, scientific machine learning, PDE-constrained optimization, certified optimization, trust-region methods, a posteriori error estimation

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_mVDDOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes SA-CPITR, a solver-assisted certified trust-region method for PDE-constrained optimization. The method uses a posteriori error bounds to translate the state and sensitivity errors of physics-informed neural surrogates into error certificates for the objective and its gradient, which govern step acceptance and termination. The authors provide analyses of certificate validity, guaranteed descent, and conditional convergence, and evaluate the method on five classes of PDE control problems and a state-constrained extension.

Critical strengths
  1. The research question is valuable. The paper goes beyond surrogate prediction accuracy to examine whether the surrogate can provide reliable information for optimization decisions. This offers an approach to addressing the exploitation of surrogate errors by the optimizer.

  2. The theoretical results support specific algorithmic decisions. The paper separately addresses certificate validity, true objective descent at accepted steps, and terminal first-order criticality. It applies the same error bounds to neural predictions and solver outputs, rather than treating numerical solver outputs as exact ground truth. These theoretical results provide an important foundation for the method.

Critical weaknesses
  1. The core theoretical guarantees have not been fully realized in the practical optimization loop. The theory requires exact integration or numerical integration with rigorous error bounds, whereas the implementation uses fixed-grid quadrature without accounting for an upper bound on the quadrature error. Consequently, the in-loop accuracy checks, acceptance ratios, and per-step descent guarantees have not been rigorously verified for the continuous PDE problem.

  2. The experiments are small in scale, leaving the method’s practical scope unclear. The six problems involve only a small number of control parameters. Although these settings are suitable for checking the certificates and algorithmic behavior, they do not adequately assess how certificate conservatism, refinement costs, and certification success rates change as control dimension and spatial complexity increase. The current experiments do not sufficiently support the method’s applicability to more complex problems.

  3. The efficiency comparison lacks matched stopping and certification criteria. The authors report runtimes that include offline training, but the comparison with the solver-in-the-loop reference optimizer does not use aligned stopping and certification criteria.

Questions for authors
  1. Can the authors perform at least one complete optimization run using numerical integration with rigorous error bounds throughout the loop, and report its effects on the certificates, step-acceptance decisions, and runtime? When the refinement budget is exhausted while the accuracy conditions remain unmet, which theoretical guarantees still apply to the implementation?

  2. With the same surrogate training, refinement procedure, and budget, what additional benefits does the full method provide over an approach that calls the solver only for terminal certification? Can a targeted controlled experiment isolate the effect of Mode A on the optimization trajectory and final result, rather than merely demonstrating tighter certificates?

  3. How do certificate conservatism, computational cost, and certification success rates change as the control dimension or the complexity of the spatial forcing increases? Can the authors provide validation beyond the current low-dimensional, structurally simple settings?

2 comments
AuthorsOpenReview ↗

Comment 5: response to Reviewer mVDD (1/2)

Thank you for the detailed review and for the concrete questions, and for finding the research question valuable and the theoretical results a foundation for the algorithmic decisions. Each question is answered with new experiments in the revision.

W1, Q1 (first part). A complete run with rigorous integration throughout the loop.

The original submission re-verified the five terminal certificates without the loop's quadrature, including a continuum-valid evaluation of the C3 network certificate at the returned point, and audited the in-loop layer against ground truth at every logged iterate (Sec. 5.2, Figure 4, Tables 8–10). New in the revision, App. F.6 reports a complete C3 run in which every certificate that enters a decision (centers, gates, trial points, terminal test and returned point) is built from cellwise interval enclosures of every integral on 2⋅1042\cdot10^4 cells and bounds the errors of the values the loop uses (Table 11).

  • Decisions. The validated run takes every decision of the grid run, namely the gate outcomes, the acceptance of both steps and the terminal stop, and returns the same control. The certified ratios of the two steps decrease by less than 0.005.
  • Certificates. The returned bound χθ+η^g\chi_\theta+\hat\eta_g is 1.89⋅10−51.89\cdot10^{-5} in the grid run and 2.12⋅10−52.12\cdot10^{-5} in the validated run, while the anchored terminal test with exact integrals gives 1.06⋅10−51.06\cdot10^{-5} at the same returned control. Most of the increase comes from the width of the cell enclosures, and the term that bounds the quadrature error of the gradient is at most 4% of the validated ηg\eta_g.
  • Runtime. 2781 s against 1971 s, i.e. 1.41 times the grid run, from 18 enclosure evaluations of about 65 s each.

The enclosures use no directed rounding, and the 2D and space–time cases would need two-dimensional cell enclosures of the network, which we did not implement. A global bound on the quadrature remainder, computed from the network weights, would not replace them, since on C3 it is about 101910^{19} times the integral it bounds, and no certificate would pass a gate or the terminal test (App. F.6). The quadrature error that the grid leaves uncontrolled is small against the certificates. For all six cases, refining the certification grid up to 16 times changes the in-loop ηg\eta_g at the SA-CPITR and CPITR controls by 0.4–6.5%, far less than the gradient certificates overestimate the true gradient errors (Table 8), and it changes no terminal decision at any of the 18 returned controls (App. F.6).

Q1 (second part). Which guarantees apply when a refinement budget is exhausted?

Theorem 3 (Guaranteed descent) uses only valid value certificates at both ends of the step and neither the gates nor the budgets, so a step accepted after a budget is exhausted with a failing gate is still a certified descent step (new sentence after Theorem 3). The terminal certificate of Theorem 4 does not depend on the history of the run. What is lost is only the progress guarantee of Theorem 5. The audit in the original submission checked all 36 accepted steps of the campaign against ground truth. The revision shows that 12 of them were computed after the budget of the gradient gate was exhausted and that all 12 decreased the true objective by at least η1predk\eta_1\mathrm{pred}_k (App. F.5). Table 9 lists for each theorem how the runs evaluate it and how it is checked.

W3 and Q2 (first part). Matched stopping criteria, and the solver only for terminal certification.

You are right that the submitted comparison with the reference optimizer was unmatched, since SLSQP was run to its own tolerance. The revision stops SLSQP on the same discretization and from the same start with the same certified test χa+ηga≤tol\chi_a+\eta_g^a\le\mathrm{tol}. It then needs 14/24/25/78/323 PDE solves on C1–C5 with sensitivity gradients and 14/18/14/24/54 with adjoint gradients, against 0/3/0/13/17 for SA-CPITR (Table 7, Figure 3a). In wall clock a single classical optimization is faster with our cheap solvers (at most 6.4 s), and Sec. 5.2 and App. F.4 state when the offline training pays off.

The CPITR arm of the submission has the same training and refinement as SA-CPITR without solves, and in the 240-run ablation SA-CPITR certifies 53 of 74 paired configurations against 26 for CPITR, at 0.72 of its wall clock (App. G). New in the revision, CPITR followed by a terminal solver check certifies all of C1–C5 with 40 PDE solves, against 33 for SA-CPITR, and it takes 2.3–5.2 times the wall clock of SA-CPITR on C2, C4 and C5, because SA-CPITR stops when its anchor certifies instead of running to the iteration cap or the futility rule. A frozen surrogate followed by the same check certifies only C1 and C3 (App. G.2, Table 16).

AuthorsOpenReview ↗

Comment 6: response to Reviewer mVDD (2/2

Q2 (second part). Isolating the effect of mode A on the trajectory and the result.

In the original submission, mode A executed in the out-of-distribution runs of App. H, where a gradient anchor sharpened ηg\eta_g by four to five orders of magnitude and one mode-A step was accepted in each run. To isolate its effect, App. H now reports paired runs on one GPU with deterministic kernels. The two runs of a pair are identical up to the first mode-A entry, and from there one uses the model of mode A while the other solves and stores the same anchor and uses it only to sharpen the certificates (Table 19).

  • Stress family (13 pairs). In two runs built with the criticality at the anchor just below tol, only the run with mode A certifies (0.9780.978 and 0.9800.980 times tol, confirmed by the exact test). The run without it stays above tol until the iteration cap, after 35 instead of 18 PDE solves. Over the family, mode A needs 160 PDE solves against 220.
  • Biased surrogate gradient (10 pairs). A surrogate trained only on the state residual has inaccurate sensitivities. The returned design differs in 9 pairs and is better with mode A in 6. On C4 the surrogate gradient is too flat or points away from descent, and mode A returns controls with J/J⋆−1J/J^\star-1 of 0.23–0.47 against 0.69–2.69. On C2 the gradient is too steep, the corrected predicted decrease falls inside the value certificates, and the run with mode A stays at the anchor.

Every accepted step is certified descent in both modes (Theorem 3), and mode A keeps the gradient gate passable through (A4) when the network's gradient certificate can no longer be refined, which Theorem 5 needs. The theory does not state which run reaches the better design within finitely many iterations, and trying the surrogate step after a rejected mode-A step, whose acceptance still certifies descent, is possible future work.

A bug we found and fixed. In the mode-A code of the submitted supplement the correction ga−gθ(ck)g_a-g_\theta(c_k) was computed once at the entry and not recomputed when the network was refined while mode A was active. All ten submitted runs with a mode-A entry were rerun with the corrected code, whose numbers Apps. G and H now report. No certification outcome changes, the anchor events fall from 25 to 21, four returned controls change and none of them gets worse, and the anchored bound at the returned control now agrees with the exact test to a relative 10−610^{-6} (up to 1.8% before).

W2, Q3. Control dimension and spatial complexity.

In every case the control is a function, for example the source term of C5 on the whole domain, and the optimization variable is its coefficient vector in a spline or modal basis, as engineering control and design problems are usually parametrized (Sec. 1). The cases of the original submission use up to 16 such coefficients (C5). App. G.3 now varies the control dimension on C2 with p=2,4,8,16p=2,4,8,16 sine modes, whose forcing frequency grows with pp, with three seeds and an exact closed-form truth (Table 17, Figure 3b).

  • Certificates. All are valid at every logged iterate, with ∣Jθ−J^∣≤0.91η^J|J_\theta-\hat J|\le0.91\hat\eta_J and ∣χθ−χ∣≤0.036η^g|\chi_\theta-\chi|\le0.036\hat\eta_g at worst.
  • Success rate. SA-CPITR certifies 3/3, 3/3, 2/3 and 0/3 seeds at p=2,4,8,16p=2,4,8,16, and CPITR certifies none. At p=16p=16 SA-CPITR still returns a true objective 229 and 1266 times below that of CPITR in two seeds.
  • Cost. 3, 5 and 10 PDE solves per certified run (9 or 10 with the larger network at p=8p=8), and a gradient anchor costs 1+p1+p solves (two with the adjoint certificate of Remark 2).
  • Conservatism. At the campaign's capacity ηg(c0)\eta_g(c_0) grows 4.1 times at p=8p=8 and 33 times at p=16p=16 relative to p=4p=4, faster than p\sqrt p. The best of eight capacities reduces this to 1.6 and 5.7 times. Run with this capacity at p=8p=8, SA-CPITR certifies all three seeds at bounds of 0.022–0.10 times tol, while CPITR certifies none (Table 17, row 8‡8^\ddagger). The limit is therefore the training of the network (A1) and not the certificate layer, which never uses the dimension of cc, i.e. the number of parameters pp.

We chose problems in 1D and 2D space and in 1D space–time on purpose, because their solutions are known in closed form or from high-accuracy Galerkin solvers, which lets us check every certificate and returned point against the truth (App. F.5). Related work uses problems of the same kind, e.g. Mowlavi & Nabi (JCP 2023) for PINN-based optimal control of the 2D Laplace and 1D Burgers equations with analytical solutions and Qian et al. (SISC 2017) for a certified reduced-basis trust region on a 2D thermal fin with two and six parameters. Our purpose is to provide the theoretical basis, which transfers to higher-dimensional controls. However, larger and three-dimensional problems need other surrogates, such as neural operators with conforming reconstructions, which the certificates cover unchanged and which we name as future work.

Reviewer_HY2QOpenReview ↗
Rating
4
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

Neural PDE surrogates accelerate PDE-constrained optimization but can lead optimizers into regions where model predictions are unreliable. To address this, the authors introduce SA-CPITR, a certified trust-region framework that converts PDE residuals into computable functional majorants to bound objective and gradient errors online. When surrogate certificates fail trust-region gates, the method selectively invokes a high-fidelity solver under a strict event budget to generate reusable "anchors" that sharpen local error bounds and supply gradient corrections. The authors theoretically prove certificate validity, guaranteed true objective descent on every accepted step, and convergence to a certified first-order critical point. Across five benchmark PDE control problems, SA-CPITR certified terminal criticality on all five using only three total anchor events (33 PDE solves), drastically outperforming uncertified frozen surrogates.

Critical strengths
  1. Rigorous Integration of Online Computable PDE Certificates into Trust-Region Optimization: The paper establishes the first framework that dynamically connects computable, functional majorant error bounds derived from neural PDE residuals directly to trust-region acceptance, step refinement, and termination criteria with theoretical convergence guarantees. Learned surrogates like PINNs often fail silently or are exploited by optimizers into regions of high prediction error. By converting state and sensitivity majorants into objective and gradient certificates, the framework guarantees true objective descent for every accepted step and certified first-order criticality upon termination without ever evaluating the true objective. This provides a principled mathematical solution to the long-standing reliability gap in neural surrogate optimization.

  2. High Computational Efficiency via Reusable Solver Anchors: The event-triggered solver policy restricts expensive numerical solves strictly to situations where computable certificates prove the surrogate's accuracy is insufficient. When called, high-fidelity solver outputs are converted into reusable "anchors" that sharpen surrogate bounds nearby or construct additively corrected trust-region models. This mechanism delivers exceptional practical utility: across five benchmark PDE control problems (C1–C5), SA-CPITR certified terminal criticality on all cases using a total of only 3 anchor events (33 solver evaluations), compared to up to thousands of solves required by full solver-in-the-loop optimizers.

Critical weaknesses

To help readers navigate the mathematical framework through the jungle of assumptions and theorems, consider adding a brief overview diagram or text outlining the logical chain of certified optimization:

  • Assumption 2 (Residual Bounding): Low PDE/tangent residuals ⇒\Rightarrow State & sensitivity errors bounded by functional majorants.
  • Theorem 1 (Assembly): State & sensitivity majorants ⇒\Rightarrow Computable certificates for objective (∣Jθ−J^∣≤ηJ\vert{}J_\theta - \hat{J}\vert{} \le \eta_J) and gradient (∥gθ−∇J^∥≤ηg\Vert{}g_\theta - \nabla \hat{J}\Vert{} \le \eta_g).
  • Theorems 3–4 (Optimization): Objective & gradient certificates ⇒\Rightarrow Guaranteed descent per step and certified terminal criticality.
Questions for authors

N.A.

1 comment
AuthorsOpenReview ↗

Comment 2: response to Reviewer HY2Q

Thank you for the very positive review and for the clear summary of our contributions. We are glad that you see the integration of computable PDE certificates into trust-region acceptance, refinement and termination as a principled answer to the reliability gap of neural surrogates in optimization. We also thank you for the concrete suggestion of an overview of the logical chain, and the revision makes this chain explicit in three places.

  1. Figure 1 now carries the labels along the proof chain, and its caption explains them. Assumption 2 and Definition 3 label the arrow from the state to the majorants, Definition 1 and Theorem 1 the majorants, Definition 2 and Theorem 2 the certificates, Theorem 3 the arrow from the trust-region step to the decision, and Theorems 3–5 the decision. Red labels mark Definition 3 and Theorem 2, which use solver outputs, in the colour of the PDE solver in the figure.
  2. Sec. 4 walks along this chain in two sentences. Small PDE and tangent residuals bound the state and sensitivity errors by majorants (Assumption 2), the assembly turns the majorants into certificates of the objective and the gradient (Theorem 1), solver anchors sharpen these certificates (Theorem 2), and valid certificates give certified descent at every accepted step and certified criticality at the returned point (Theorems 3 and 4).
  3. A proof map at the start of App. B (Figure 5) draws the same chain from the PDE residuals to convergence. Each arrow carries its assumption or theorem together with the hypotheses and earlier statements that its proof uses and the appendix that contains the proof. For the experiments, Table 9 (App. F.5) now also names, for each check against ground truth, the theorem it tests.
Reviewer_CjAdOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces a trust-region method, SA-CPITR, for PDE-constrained optimization with a parametric physics-informed neural surrogate. The method uses residual-based a posteriori majorants to bound errors in the surrogate state and sensitivities. It then converts these bounds into certificates for the reduced objective and gradient. The paper proves certificate validity, guaranteed descent for accepted steps, and conditional convergence to first-order criticality. In the main experiments, the authors show that SA-CPITR certifies terminal criticality on five PDE control problems.

Overall, the central idea appears interesting and the approach seems novel. However, the gap between the proved theorems and the implemented experiments should be clarified, and the practical advantages over standard PDE optimization methods remain somewhat unclear.

Critical strengths
  • The problem being addressed is important.
  • The authors presented a coherent framework. The theorems appear useful and are presented in an understandable fashion.
  • The numerical examples are clearly presented.
Critical weaknesses
  • The gap between the proved theorems and the experiments should be clarified.
  • Practical advantages seem a bit unclear.
Questions for authors
  • Which theoretical guarantees apply to the implemented experiments, given the approximate quadrature and finite refinement budgets?
  • What practical advantage does SA-CPITR offer over a standard adjoint-based optimizer or surrogate optimization followed by terminal verification?
  • How do total wall time and actual PDE-solve counts compare with a standard adjoint-based optimizer and with a certified reduced-basis or adaptive multifidelity method?
  • Can Proposition 1 be reformulated using only the realized-step Taylor remainder actually needed by the proof, without claiming that (17) establishes P1 on the trust region?
2 comments
AuthorsOpenReview ↗

Comment 3: response to Reviewer CjAd (1/2)

Thank you for the careful review and for finding the framework coherent and the theorems useful and understandable. We address the two weaknesses through your four questions.

Q1. Which guarantees apply to the experiments, given quadrature and finite refinement budgets?

In the implemented runs, the certified criticality of Theorem 4 is confirmed at all five returned points of C1–C5, whose certificates are re-verified without the loop's quadrature. Theorem 3 (Guaranteed descent) needs valid certificates inside the loop, which the loop evaluates on a grid, and the audit against ground truth confirms them at every logged iterate. The finite budgets affect neither theorem, and only the convergence of Theorem 5 is not exercised by runs of at most ten iterations, which Sec. 4 states.

The original submission separated the returned points, whose certificates were re-verified without the loop's quadrature, from the in-loop layer, which was audited against ground truth (Sec. 5.2, Figure 4, Tables 8–10). The revision adds a continuum-valid run and a statement on finite budgets, and Table 9 (App. F.5) now lists both layers per theorem.

  • Returned points (original submission). The certificate of Theorem 4 at the returned point does not depend on the history of the run. All five terminal certificates of C1–C5 were checked again without the loop's quadrature, by exact integration of the anchored certificates on C2, C4 and C5, by solver anchors placed after the run on C1 and C3, and on C3 also by a continuum-valid evaluation of the network certificate, and every check stays below tol (App. F.6).
  • In-loop layer. Gates, acceptance ratio and descent are evaluated on the grid. The audit in the original submission found the objective certificates valid at all 71 logged centers and the criticality transfer at all 53, true descent of at least η1predk\eta_1\mathrm{pred}_k at all 36 accepted steps and no false pass of the gradient gate (Figure 4, Table 8, App. F.5). New in the revision, a complete C3 run in which every certificate that enters a decision is continuum-valid makes the same decisions and returns the same control as the grid run, at 1.41 times the wall clock (App. F.6, Table 11). Refining the certification grid up to 16 times changes no terminal decision at any of the 18 returned controls.
  • Finite budgets (new). Theorem 3 (Guaranteed descent) uses only valid value certificates and neither the gates nor the refinement budgets (new sentence after Theorem 3). A step accepted after a budget has run out is therefore still a certified descent step, and only the progress guarantee of Theorem 5 is lost. Of the 36 audited steps, 12 were of this kind, and all 12 decreased the true objective by at least η1predk\eta_1\mathrm{pred}_k (App. F.5).

Sec. 3.1 adds that every numerical PDE optimizer carries discretization and quadrature errors in its objective and gradient, whereas here the majorants bound the distance to the exact PDE solution and the quadrature of their integrals is the only approximation in a certificate.

Q4. Proposition 1 with the realized-step remainder.

Yes, thank you for pointing this out. Property (P1) of Definition 4 is now the bound ∣Rkm∣≤κH2∥sk∥2|R^m_k|\le\tfrac{\kappa_H}{2}\Vert s_k\Vert^2 on the model's Taylor remainder along the computed step, which (A2) implies. The definitions (17) turn this bound into an equality, so Proposition 1(i) holds as stated without any claim about the Hessian on the trust region. Lemmas 3 and 4 and the proof of Theorem 5 use (P1) only in this form (App. C.4).

AuthorsOpenReview ↗

Comment 4: response to Reviewer CjAd (2/2)

Q2. Practical advantage over an adjoint-based optimizer or surrogate optimization followed by terminal verification.

The submission compared SA-CPITR with a frozen surrogate and with CPITR, which uses the same training and refinement without solves. The certified bound of the frozen surrogate exceeds tol by a factor of about 10 to 270 on four of the five cases (Table 1, Figure 2a), and in the 240-run ablation SA-CPITR certifies 53 of 74 paired configurations against 26 for CPITR, with 27 pairs decided in its favour and none against (App. G). The revision adds the following comparisons.

  • Surrogate, then terminal verification (App. G.2, Table 16). A solver check after the run, which solves for the state and its pp sensitivities at the returned control and applies the terminal test of Theorem 4, certifies only C1 and C3 for the frozen surrogate (40 PDE solves). Repairing the three failed designs classically costs 474 further solves with sensitivity gradients or 102 with adjoint gradients. After CPITR the same check certifies all of C1–C5 with 40 PDE solves, against 33 for SA-CPITR, and CPITR takes 2.3–5.2 times the wall clock of SA-CPITR on C2, C4 and C5, because SA-CPITR stops as soon as its anchor certifies.
  • Classical optimizer with the same certified stop (App. F.4, Table 7, Figure 3a). SLSQP on the same discretization, from the same start and stopped by the same test χa+ηga≤tol\chi_a+\eta_g^a\le\mathrm{tol} needs 14/24/25/78/323 PDE solves on C1–C5 with sensitivity gradients and 14/18/14/24/54 with adjoint gradients, against 0/3/0/13/17 for SA-CPITR. The adjoint count is a count only, since none of our solvers has an adjoint, and it favours the classical optimizer, because our own anchors are counted with sensitivities.
  • Certificates outside the loop. The certificates bound the errors of a trained surrogate at any control, so they also apply when the surrogate is used without our trust-region loop. The certified bound of the frozen surrogate in Table 1 is such a use (contribution (i) now says this).

Q3. Wall time and PDE solves against an adjoint optimizer and against reduced-basis or multifidelity methods.

The submission reported the run times of SA-CPITR including offline training (3–56 minutes per case) and the solve counts of the reference optimizer run to its own tolerance (App. F.4). Our reference solvers are cheap (0.01–18 ms per solve), so a single classical optimization to the certified stop takes at most 6.4 s. We now state this in Sec. 5.2. The offline training pays off for a single optimization once one solve costs more than 6–62 s (C1–C5, both gradient conventions), and earlier when one training serves several optimizations on the same parametrized PDE, since the surrogate is trained on residuals only and does not depend on the objective (App. F.4). Certified reduced-basis trust-region methods show the same pattern, e.g. Qian et al. (SISC 2017) report 14–39% of the full-order solves at 44–83% of the full-order wall time.

We did not implement a reduced-basis or multifidelity comparator. These methods need intrusive access to the full-order operators, offline full-order snapshots and an affine or empirically interpolated parameter dependence, and their bounds measure the distance to the full-order model, whereas our majorants bound the distance to the PDE solution (App. I). The two approaches are complementary, and since the majorant certifies any conforming candidate, a reduced-basis approximation could itself be the certified candidate in SA-CPITR.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Comment 1: general response (to all reviewers and the AC)

Summary of the revision

We thank the three reviewers for their careful reading and their constructive comments. The reviewers find the problem important (CjAd, mVDD), the framework coherent with theorems that support the algorithmic decisions (CjAd, mVDD), and the combination of online certificates with trust-region acceptance and reusable solver anchors rigorous and efficient (HY2Q). For each point below we name what the submission already contained and what the revision adds.

  1. Logical chain of the theory (HY2Q). Figure 1 now labels each stage with the assumption, definition or theorem it uses, Sec. 4 walks along this chain, and a proof map at the start of App. B (Figure 5) gives the hypotheses of every proof and where the proof is.
  2. Which guarantees hold in the experiments (CjAd Q1, mVDD W1, Q1). The submission already separated the returned points, whose certificates were re-verified without the loop's quadrature, from the in-loop layer, which was audited against ground truth at every logged iterate (Sec. 5.2, Figure 4, Tables 8–10, Apps. F.5 and F.6). The revision adds a theorem column to Table 9 and states that Theorem 3 (Guaranteed descent) needs neither the gates nor the refinement budgets (Sec. 4). All 12 steps accepted after a refinement budget ran out decreased the true objective (App. F.5), and a complete C3 run with continuum-valid certificates in every decision makes the same decisions and returns the same control as the grid run, at 1.41 times the wall clock (App. F.6, Table 11).
  3. Costs and baselines (CjAd Q2, Q3, mVDD W3, Q2). The submission compared SA-CPITR with CPITR and a frozen surrogate on all cases and in a 240-run ablation, and it reported the solve counts of the solver-in-the-loop reference optimizer, which ran to its own tolerance. The revision stops this optimizer (SLSQP) with the same certified test as SA-CPITR, and it then needs 14/24/25/78/323 PDE solves on C1–C5 with sensitivity gradients and 14/18/14/24/54 with adjoint gradients, against 0/3/0/13/17 for SA-CPITR (Figure 3a, App. F.4). The revision also adds a frozen surrogate or CPITR followed by a terminal solver check (App. G.2) and states the wall-clock break-even (Sec. 5.2, App. F.4).
  4. Control dimension (mVDD W2, Q3). In every case the control is a function, for example the source term of C5 on the whole domain, and the optimization variable is its coefficient vector in a spline or modal basis, as engineering control and design problems are usually parametrized (Sec. 1). The cases of the submission use up to 16 such coefficients. The revision adds a sweep on C2 with p=2,4,8,16p=2,4,8,16 sine modes, whose forcing frequency grows with pp, in which SA-CPITR certifies all seeds up to p=8p=8 once the network capacity is large enough and CPITR certifies none (Sec. 5.2, Figure 3b, App. G.3).
  5. Mode A (mVDD Q2). In the submission, mode A executed only in the out-of-distribution and policy-study runs (Apps. G and H). The revision adds paired runs that are identical up to the first mode-A entry and then continue with and without mode A (App. H, Table 19).
  6. Proposition 1 (CjAd Q4) is stated with the realized-step remainder only (App. C.4).

While setting up the paired runs of item 5 we found and fixed a bug in the mode-A code of the submitted supplement (details in our reply to mVDD). Mode A never executes in the campaign or in the 240-run ablation, so these results are not affected, and the ten affected runs were rerun with the corrected code. The revised supplement contains the fix and the code, logs and scripts of all new experiments.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.