- 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
-
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.
-
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
-
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.
-
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.
-
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
-
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?
-
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?
-
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
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 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 ( and 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 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 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 (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 sine modes, whose forcing frequency grows with , with three seeds and an exact closed-form truth (Table 17, Figure 3b).
- Certificates. All are valid at every logged iterate, with and at worst.
- Success rate. SA-CPITR certifies 3/3, 3/3, 2/3 and 0/3 seeds at , and CPITR certifies none. At 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 ), and a gradient anchor costs solves (two with the adjoint certificate of Remark 2).
- Conservatism. At the campaign's capacity grows 4.1 times at and 33 times at relative to , faster than . The best of eight capacities reduces this to 1.6 and 5.7 times. Run with this capacity at , SA-CPITR certifies all three seeds at bounds of 0.022–0.10 times tol, while CPITR certifies none (Table 17, row ). The limit is therefore the training of the network (A1) and not the certificate layer, which never uses the dimension of , i.e. the number of parameters .
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.
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⋅104 cells and bounds the errors of the values the loop uses (Table 11).
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 1019 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 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 (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. 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).