acceptodds
Under review as a conference paper at ICLR 2027

Learning PDE-Feasible Design Manifolds for Solver-Amortized Optimization

Abstract

PDE-constrained optimization repeatedly solves a design-dependent state equation. For a class of separable elliptic PDEs, we show that one reference solution induces coefficient–state pairs that preserve compatible source and boundary data and satisfy the residual-inheritance identity . An exact reference yields a family of exact states, while an approximate reference transfers an explicit residual certificate. This enables hard-constrained design updates from cached fields without a new state solve. Because a fixed family spans only restricted directions, a full-space check at restricted stationarity constructs a bound- and volume-constrained Wendland-RBF candidate from the inaccessible gradient component; one candidate solve and an acceptance gate decide whether to update the reference. Across thermal experiments, analytical updates take s, versus s for independent PINN re-solves. In end-to-end physics-informed optimization, our method trains state PINNs rather than and reaches a lower FVM-audited resistance. The framework therefore amortizes expensive state solves through residual-certified reuse.

Primary area:
neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)
Keywords:
PDE-constrained optimization, scientific machine learning, physics-informed neural networks, solver-amortized optimization, PDE-feasible design manifolds, topology optimization

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_hkskOpenReview ↗
Rating
3(+1)
Confidence
1(−1)
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a method to reduce repeated solves in PDE-constrained optimization. The authors use a reference PDE solution to construct feasible designs and states, and perform multiple optimization updates within a restricted design space without re-solving the PDE. When the restricted optimization can no longer make descent progress, the algorithm updates the reference solution through gradient diagnostics and RBF corrections.

Critical strengths
  1. The authors leverage the analytical structure of the PDE to reuse reference solutions, thereby reducing repeated solves. This idea has a clear theoretical basis.
  2. The experiments demonstrate the potential to reduce solver calls
Critical weaknesses

The main objective of this paper is to reduce computational costs by amortizing PDE solves. However, Section 4.2 only compares the number of solver calls between Ours + FVM and Baseline + FVM without reporting the total wall-clock runtime, making it difficult to assess the practical utility of the proposed algorithm. The authors should either demonstrate an end-to-end speedup of Ours + FVM over Baseline + FVM under certain settings, or show an end-to-end speedup of Ours + PINN over Baseline + PINN in settings where PINNs inherently offer advantages over conventional FVM solvers.

Questions for authors
  1. Appendix C.2 specifies that candidate PINN retraining uses at most 1,200 Adam steps, yet Appendix C.4 states that 11 of the 17 candidate retrainings in the baseline reached a 200-step cap without passing the held-out state-accuracy gate. What is the rationale behind this 200-step cap? If the training budget for the baseline were increased, would it reduce the rejection rate and lead to better optimization outcomes?
  2. Can the authors prove that a non-zero accessible component necessarily corresponds to a descent direction within the current design family? For example, consider the simplex constraint ρ1+ρ2+ρ3=1\rho_1 + \rho_2 + \rho_3 = 1, the current iterate ρ=(0,0,1)\rho = (0, 0, 1), a feasible design family ρ(t)=(t,t,1−2t)\rho(t) = (t, t, 1 - 2t) (with 0≤t≤1/20 \le t \le 1/2), and the objective j(ρ)=ρ1−ρ2j(\rho) = \rho_1 - \rho_2. The objective is identically zero across the entire design family (meaning no descent direction exists within the family). However, with W=IW = I and γg=1\gamma_g = 1, Eq. (22) still yields a non-zero accessible component (gρ,∥=(−1/4,−1/4,1/2)T≠0g_{\rho, \parallel} = (-1/4, -1/4, 1/2)^T \ne 0).
  3. Could the authors provide the computational complexity of the non-PDE components, such as solving the QP subproblem in Eq. (23) and the RBF construction? This would help substantiate the scalability of the proposed framework.
4 comments
AuthorsOpenReview ↗

Demonstration of end-to-end wall-clock speedup under a shared adaptive PINN retraining protocol

We thank the reviewer for emphasizing that a reduction in state-solver calls does not by itself establish a practical wall-clock advantage. We address the reviewer’s second suggested comparison by evaluating the proposed and every-update methods with the same PINN-based state solver under a common adaptive retraining protocol. Both methods use the same initial design, pretrained reference network, maximum of 1,200 Adam steps per candidate retraining, held-out state-accuracy gates, hardware, and independent FVM evaluation. The reported algorithm time includes all optimization costs: cheap-phase updates, full-space diagnostics, QP solution, RBF construction, analytical distillation, state-network retraining, data transfer, and bookkeeping. Post-hoc FVM audits are excluded from both timings because they do not affect either optimization trajectory.

Table 1. End-to-end PINN comparison under the common adaptive retraining protocol.

MethodState solvesAlgorithm time (s)FVM-audited JJ
Every-update PINN16161683.81683.81.54291.5429
Proposed22251.1251.11.06851.0685

Under this common protocol, the proposed method achieves a 30.7530.75\\% lower independently audited objective while reducing algorithm wall-clock time by 85.0985.09\\%. The independently FVM-audited quality-versus-time curve further shows that the every-update baseline does not reach the proposed method’s audited design quality anywhere along its observed trajectory; hence, no equal-quality baseline time exists within the recorded run. The shared reference-network pretraining requires 185.55185.55 s and is excluded from both algorithm times because the same pretrained reference is used by both methods. Even when this shared cost is included, the corresponding end-to-end totals are 436.7436.7 s for the proposed method and 1869.41869.4 s for the every-update baseline, so the conclusion is unchanged. We will add the timing protocol, complete cost accounting, comparison table, and independently audited quality-versus-time curve to the revised manuscript.

AuthorsOpenReview ↗

Addressing Question 1: Correction of the mismatched retraining-step budget and controlled comparison

We thank the reviewer for identifying this inconsistency. The 200-step cap was an implementation-specific budget used in the original every-update baseline, whereas the proposed refresh retraining followed the 1,200-step maximum described in Appendix C.2. We agree that this mismatch made the original timing, rejection-rate, and design-quality comparison difficult to interpret. We therefore no longer use the original comparison as evidence of computational superiority. We reran both methods under the same adaptive retraining protocol. Every candidate is evaluated every 20 Adam steps, may terminate early only after passing the same held-out state-accuracy gates, and otherwise receives a common maximum of 1,200 steps. There is no failure-based early rejection before this common limit.

Table 2. Effect of the common 1,200-step adaptive retraining protocol.

MethodTrials/passedState solvesTime (s)FVM-audited JJ
Every-update, original 200-step cap17/617/61818373.5373.51.57341.5734
Every-update, common 1,200-step cap15/1015/1016161683.81683.81.54291.5429
Proposed, common 1,200-step cap1/11/1 refresh22251.1251.11.06851.0685

Increasing the baseline budget substantially reduces its rejection rate: 10 of 15 candidates pass, compared with 6 of 17 under the original cap. It also improves the independently audited FVM objective from 1.57341.5734 to 1.54291.5429, but the improvement is only 1.941.94\\%, while runtime increases from 373.5373.5 s to 1683.81683.8 s. Thus, the reviewer’s concern is valid: part of the original rejection rate was caused by the restrictive 200-step budget. However, correcting this mismatch does not change the principal conclusion. Under the common 1,200-step protocol, the proposed method still achieves a 30.7530.75\\% lower independently audited objective while using 85.0985.09\\% less algorithm time than the retrained every-update baseline. We will revise Appendices C.2 and C.4 to state one consistent training protocol and replace the original comparison with these results.

AuthorsOpenReview ↗

Addressing Question 2: Clarification of the gradient-accessibility diagnostic and descent condition

We thank the reviewer for this counterexample. We agree that gρ,∥≠0g_{\rho,\parallel} \neq 0 does not necessarily guarantee an in-family descent direction. Indeed, in the given example,

∇jTdρdt=(1,−1,0)T(1,1,−2)=0\nabla j^T \frac{d\rho}{dt} = (1,-1,0)^T (1,1,-2) = 0,

so the objective is constant along the family. We will revise the text surrounding Eq. (22) to remove the claim that a non-zero accessible component guarantees descent and instead define gρ,∥g_{\rho,\parallel} and gρ,⊥g_{\rho,\perp} as diagnostic measures of gradient accessibility and mismatch. The correct first-order condition is δF<0\delta_F < 0, where δF=min⁡∇ρjTJθdθ\delta_F = \min \nabla_\rho j^T J_\theta d_\theta with the constraints dθ∈Tθd_\theta \in T_\theta and ∥dθ∥≤1\|d_\theta\| \le 1. Here Jθ=∂ρ/∂θJ_\theta = \partial\rho/\partial\theta, and TθT_\theta denotes the feasible parameter-direction set. Thus, Eq. (22) is a diagnostic and is not equivalent to this descent condition. This clarification does not change the algorithm or experiments: actual descent is determined from evaluated feasible updates, while Eq. (22) is used only to trigger a refresh attempt, which remains subject to the feasibility and acceptance checks.

AuthorsOpenReview ↗

Addressing Question 3: Complexity and timing of the RBF-QP refresh step

We thank the reviewer for requesting this clarification. Let NN be the number of active design cells and MM the number of multiscale RBF weights. Constructing Φ∈RN×M\Phi \in \mathbb{R}^{N\times M} requires O(NM)O(NM) time and storage in the current dense implementation. Forming ΦTWΦ\Phi^T W \Phi requires O(NM2)O(NM^2) time and O(M2)O(M^2) additional storage. The subsequent QP has MM variables, one volume equality, and 2N2N cellwise bounds. Its exact solution cost is solver- and active-set-dependent; the dense linear algebra is governed by MM, while evaluating the linear constraints scales as O(NM)O(NM) per iteration. Because M≪NM \ll N, the QP is not solved in the full NN-dimensional design space. The compact support of the Wendland basis also permits a sparse implementation for larger meshes.

Table 3: Observed non-PDE refresh overhead in the reported eight-terminal PINN experiment.

CaseActive cells NNCentersWeights MMRBF+QP (s)Total (s)
Eight-terminal PINN3784288426.37251.12

The three support radii produce M=3×28=84M = 3\times 28 = 84 weights. The measured RBF-construction and QP stage accounts for 10.50% of the complete algorithm time; the state-retaining stage remains the dominant cost. The current logs record RBF construction and QP solution jointly, so we report their combined measured time rather than introducing an unsupported split. We will add the complexity statement, implementation detail, and timing table to the revised appendix.

Reviewer_Df5QOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a solver-amortized framework for PDE-constrained design optimization by constructing analytically feasible design–state families from a single reference solution. For a class of separable elliptic PDEs, the authors derive a coupled transformation of the material coefficient and the PDE state that preserves the governing equation under compatible source and boundary conditions. In particular, the transformed PDE residual satisfies the identity (R_\theta=c_\theta R_0), allowing multiple design updates without repeatedly solving the state equation.

Critical strengths

The most appealing aspect of this work is the analytical coupling between the material coefficient and the PDE state. Rather than approximating the solution operator or repeatedly enforcing the PDE through a soft physics-informed loss, the authors exploit the mathematical structure of the governing equation to construct new design–state pairs with inherited PDE residuals.

Critical weaknesses

The analytical construction relies on the separable coefficient structure (k(\rho,u)=a(\rho)b(u)) and specific compatibility conditions on the source and boundary data. More importantly, the transformed flux satisfies (a_\theta\nabla K_\theta=c_\theta a_0\nabla K_0). Consequently, the flux directions and spatial flow paths cannot change within a fixed reference-induced family. This imposes a strong restriction on the accessible design space, particularly for topology optimization problems where changes in flow paths may be essential. While the proposed reference refresh mechanism partly addresses this limitation, the overall effectiveness may depend strongly on the reference initialization and refresh strategy. A more systematic analysis of this dependence would strengthen the paper.

The reduction in state-solver calls is convincing, but this does not always translate into lower total runtime. For example, in the end-to-end physics-informed experiment, the reported wall-clock time is 562 seconds for the proposed method versus 373 seconds for the every-update PINN baseline. The authors explain that the baseline terminates early because multiple candidate solutions fail the accuracy checks. Nevertheless, this makes the practical efficiency comparison less straightforward. Additional experiments comparing objective quality under matched wall-clock budgets would provide a more complete assessment of the computational benefits. Furthermore, the four primary thermal topology benchmarks are relatively small and symmetry-constrained. Comparisons against stronger conventional PDE-constrained optimization approaches, or more challenging design problems, would help establish the practical generality of the approach.

Questions for authors

see above

1 comment
AuthorsOpenReview ↗

Characterization of reference-family limitations, initialization sensitivity, and solver amortization

We thank the reviewer for identifying these limitations. We will state more explicitly that the analytical construction applies to the specified separable elliptic PDE class and compatible data. We also agree that a fixed reference family preserves flux directions; reference refreshing is therefore essential when progress requires flow-path reorientation.

We added a seven-source/one-sink case with only a single mirror-symmetry constraint, designed to require such reorientation.

The first cheap phase reduces JJ from 434.33434.33 to 254.37254.37 while changing the mean flux direction by only 0.035∘0.035^\circ. After three accepted refreshes, the mean direction change is 8.90∘8.90^\circ, 32.232.2\\% of active cells rotate by more than 15∘15^\circ, and JJ reaches 136.91136.91. This directly shows both when the fixed-family restriction becomes consequential and how refreshes overcome it.

We further tested the reference-initialization dependence using volume-preserving vertical perturbations of amplitudes 0.010.01, 0.020.02, and 0.030.03, with all other settings fixed.

Table 1. Reference-initialization sensitivity in the reduced-symmetry case.

Initial referenceTerminal continuous JJAccepted refreshesState solves
Uniform136.912136.9123377
Top-biased, 0.010.01124.524124.5242266
Bottom-biased, 0.010.01145.093145.0932266
Top-biased, 0.020.02119.892119.8922266
Bottom-biased, 0.020.02109.879109.8793377
Top-biased, 0.030.03121.739121.7392266
Bottom-biased, 0.030.03144.052144.0522266

The worst terminal objective is only 5.985.98\\% above the uniform result, four of the six perturbed runs improve upon it, and every run uses two or three accepted refreshes and six or seven state solves. Thus, initialization affects the optimization path but does not cause a large deterioration under the tested perturbations. For the refresh strategy, the diagnostic-based method gives a binary-audited objective of 141.01141.01, compared with 143.73143.73, 157.87157.87, 217.76217.76, and 179.07179.07 for periodic intervals of 6060, 120120, 240240, and 300300, respectively. These runs use different optimization state-solve budgets (7 for the diagnostic, 12 for Periodic-60, and 5 for each remaining schedule), so this is a sensitivity analysis rather than a matched-budget comparison. We will report both analyses and qualify the remaining dependence.

We replaced the unequal-budget PINN comparison with a common adaptive protocol: both methods use the same 1,200-step maximum and accuracy gates. The proposed method reaches an FVM-audited J=1.0685J=1.0685 in 251.1251.1 s, whereas the baseline finishes at J=1.5429J=1.5429 in 1683.81683.8 s. At the matched 251.1251.1 s budget, the baseline's best audited value is 1.68741.6874 (at 183.4183.4 s), so the proposed value is 36.6836.68\\% lower. The baseline never reaches the proposed quality on the recorded quality--time curve.

Regarding the request for comparisons against stronger conventional PDE-constrained optimization approaches, we note that the every-step FVM baseline in Section 4.2 is itself a conventional SIMP-style PDE-constrained optimizer---it re-solves the state after every design update using the same SIMP penalty and gradient-based update. The proposed framework's advantage is not a faster constrained-optimization algorithm but a solver-amortization layer that is orthogonal to the inner optimizer: the cheap phase and refresh mechanism reduce state-solver calls regardless of whether the design update uses Adam, an optimality criterion (OC), or MMA. Any such optimizer may serve as the cheap-phase engine inside Algorithm 1 without changing the amortization structure. We have added this clarification to the revision.

We will add these flow-reorientation, initialization, refresh-strategy, and matched-time results to the revision and explicitly qualify the method's PDE scope and remaining reference dependence.

Reviewer_pMp6OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies how to reduce repeated state solves in PDE-constrained material optimization. For a class of separable scalar elliptic PDEs, the authors jointly transform the material coefficient and state of a reference design to construct a family of design–state pairs with explicit flux and residual inheritance relations. Under compatible source and boundary conditions, this construction enables multiple restricted design updates using cached reference fields without additional state solves. To explore directions outside the current family, the method combines a full-space gradient diagnostic, constrained RBF corrections, and candidate-state checks to determine whether to refresh the reference. Thermal experiments using numerical and PINN-based state solvers report fewer state solves with comparable or better final design quality.

Critical strengths
  • The paper offers a meaningful optimization perspective on repeated state solves. By exploiting the governing equation’s structure, the method jointly constructs changes in the design and state so that some design updates can directly reuse an existing physical field. Connecting computational savings to the choice of search space provides a valuable direction for PDE-constrained optimization.

  • The continuous analytical construction is transparent and its scope is explicitly stated. Theorem 1 and Eqs. (12)–(16) provide directly verifiable flux and residual relations. The presentation distinguishes exact from approximate reference states and specifies compatibility requirements for different source and boundary conditions. The authors also acknowledge the limited expressiveness of a fixed reference family, helping readers understand both the guarantees and their conditions.

  • The analytical reuse mechanism receives direct numerical validation. Section 4.1 and Appendix A evaluate transformed states against independent FVM solutions and compare linear, geometric, and harmonic paths at matched endpoint coefficient contrast. The observed error trends are consistent with the transformation-sensitivity analysis, providing useful evidence for the feasibility of reference-field reuse and its error-propagation mechanism.

  • The end-to-end results demonstrate practical potential and include useful verification. Across the four reported thermal benchmarks, the aggregate number of full state solves decreases from 155 to 42, with final objectives no worse than those of the every-step-solve baseline. The PINN experiment additionally evaluates final designs using an FVM solver independent of network training. Counting rejected candidates and reporting unsuccessful refreshes and timing breakdowns make the method’s practical behavior and limitations easier to assess.

Critical weaknesses
  • The numerical interpretation of the theoretical guarantees needs clarification. Theorem 1 and Eqs. (12)–(16) rely on the continuous chain rule, whereas Appendix B uses a finite-volume discretization with harmonic face coefficients. The continuous residual identity does not automatically imply an identical relation for the discrete residuals of the transformed grid states. Please explain how new states are constructed from an FVM reference solution, specify how residual certification is evaluated in the implementation, and report discrete residuals or conservation errors at representative cheap-phase iterates. Moreover, Eq. (10) defines an admissible parameter set without sufficiently explaining how Adam updates of the general modulation network maintain both density bounds and the volume constraint. Appendix A provides normalization for prescribed transformation paths; the corresponding procedure for the full optimizer, together with refresh thresholds and QP regularization parameters, would help readers understand and reproduce the practical guarantees.

  • The trade-off between the restricted design space and reference refreshing requires more targeted validation. Equation (12) preserves the flux directions and relative spatial distribution within a fixed reference family. Locations with the same reference potential must also share the same relative coefficient modulation, limiting the ability to reconstruct heat-flow paths. The authors explicitly acknowledge this restriction and introduce reference refreshing; Figure 4 also shows objective reductions during both cheap phases and refreshes. However, these trajectories do not isolate when the restriction becomes consequential, whether a small number of refreshes adequately compensates for it, or how much benefit comes from analytical reuse versus RBF corrections. A targeted example within the supported PDE class that requires substantial flow-path reorientation would be informative. Under matched state-solve budgets, the full method could be compared with a baseline using the same RBF representation and constraints but full-space-gradient-driven updates without analytical cheap phases. A comparison with periodic refreshing would further clarify the value of the missing-gradient diagnostic.

  • The practical trade-off between computational cost and design quality remains unclear. Table 4 reports 3 versus 18 state-network trainings, but optimization times of 562 versus 373 seconds, with better final design quality for the proposed method. These timings exclude shared reference pretraining and post-hoc FVM audits. The results support a reduction in training calls, but do not establish the computational advantage at comparable design quality. Furthermore, 11 of the baseline’s 17 candidate updates fail the state-accuracy gate, and the run terminates through step-size reduction rather than demonstrated design stationarity. Its poorer outcome may therefore also depend on the training budget and rejection rules. Please clarify the relationship between the maximum of 1200 Adam steps in Section C.2 and the 200-step cap in Section C.4, including any early-rejection mechanism and whether it applies consistently to both methods. Using the stated common accuracy gates, independently FVM-audited design-quality-versus-time curves would make the practical benefit easier to assess.

  • Inconsistencies in Appendix D prevent a consistent interpretation of its quantitative results. Figure 8(d) reports final resistances of 0.104 for the proposed method and 0.096 for the baseline, whereas Table 5 reverses these values, changing the performance ranking. The proposed trajectory also ends near horizontal coordinate 16, while Table 5 reports 9 state solves. Please clarify whether the figure’s “cumulative FVM state solves” and the table use the same counting convention, including whether additional independent FVM audits are counted in the figure. The figure, table, and text should be reconciled against the original run logs; if they correspond to different evaluation settings, those differences should be stated explicitly. Until clarified, the reported performance ranking and the claimed 79.5% reduction in solves cannot be jointly verified from this experiment.

Questions for authors

Please refer to weaknesses.

6 comments
AuthorsOpenReview ↗

Clarification on the numerical interpretation of theoretical guarantees

We thank the reviewer for identifying an important distinction between the continuous construction and its finite-volume realization. We agree that Theorem 1 and Eqs. (12)–(16) establish an exact identity for the continuous PDE, but do not imply an exact algebraic identity for a finite-volume discretization with harmonic face interpolation. We will revise the manuscript to state this distinction explicitly and will no longer describe the transformed grid state as an exact solution of the discrete FVM system.

Construction from an FVM reference state. For the four FVM-based benchmarks in Section 4.2 and Appendix B, each reference state is obtained by solving

Ah(ρ0)K0,h=bhA_h(\rho_0)K_{0,h}=b_h

with the cell-centered FVM and harmonic face coefficients. The resulting cell-centered potential K0,h,iK_{0,h,i} is cached. During a cheap phase, the transformed coefficient, density, and state are constructed pointwise as

\rho_{\theta,i}=a^{-1}(a_{\theta,i}),\qquad K_{\theta,h,i}=F_\theta(K_{0,h,i}).$$ Evaluating $K_{\theta,h}$ requires only network evaluation and one-dimensional quadrature for $F_\theta$; no transformed FVM system is solved during the cheap updates. At each cheap-phase endpoint, the harmonic face coefficients are recomputed from $a_{\theta,i}$, and the transformed FVM operator is assembled. We then evaluate $$\epsilon_R= \frac{\\|A_h(\rho_\theta)K_{\theta,h}-b_h\\|_2} {\\|b_h\\|_2}.$$ The full-space check at the same endpoint independently solves the FVM system on the transformed design, allowing us also to report $$\epsilon_K= \frac{\Vert K_{\theta,h}-K_{\theta,h}^{\mathrm{FVM}} \Vert_2}{\Vert K_{\theta,h}^{\mathrm{FVM}} \Vert_2}.$$ If a refresh is accepted, its independently solved FVM state becomes the reference state for the next cheap phase. **Table 1.** Discrete certification metrics at representative cheap-phase endpoints. The worst residual/error across the sequence is also reported. | Case | Checkpoint | $\epsilon_R$ | $\epsilon_K$ | | :--- | :--- | :--- | :--- | | Example 1 | First | $9.21\times10^{-4}$ | $1.95\times10^{-4}$ | | | Last | $3.63\times10^{-11}$ | $1.67\times10^{-12}$ | | | Worst | $1.44\times10^{-3}$ | $7.47\times10^{-4}$ | | Example 2 | First | $7.36\times10^{-2}$ | $1.46\times10^{-3}$ | | | Last | $9.02\times10^{-12}$ | $2.03\times10^{-12}$ | | | Worst | $1.28\times10^{-1}$ | $4.48\times10^{-3}$ | | Example 3 | First | $6.25\times10^{-3}$ | $9.81\times10^{-4}$ | | | Last | $2.73\times10^{-10}$ | $9.56\times10^{-12}$ | | | Worst | $7.74\times10^{-2}$ | $1.27\times10^{-3}$ | | Example 4 | First | $2.87\times10^{-3}$ | $1.56\times10^{-3}$ | | | Last | $1.05\times10^{-11}$ | $2.86\times10^{-12}$ | | | Worst | $1.56\times10^{-1}$ | $2.81\times10^{-3}$ | Although the algebraic residual is sensitive to local face-flux discrepancies, the independently verified state discrepancy remains below $4.48\times10^{-3}$ at all reported checkpoints. **Enforcement of the admissible set.** We agree that Eq. (10) did not sufficiently describe how feasibility is maintained during Adam updates. Adam updates unconstrained network weights, but every forward evaluation maps those weights to a feasible design. Specifically, $$h_\theta(s)=\delta\tanh(\widetilde h_\theta(s)),\qquad \delta=0.55.$$ A scalar log-ratio bias is found by monotone bisection over the interval permitted by the coefficient bounds so that the prescribed mean density is satisfied. If the complete modulation shape admits no feasible bias, it is scaled along the ray from the identity transformation to the largest feasible value. An implicit-function correction supplies the derivative of the volume-normalizing bias, so the optimization gradient remains tangent to the hard volume constraint. Feasibility is therefore enforced by the parameter-to-design map at every forward evaluation rather than by post-hoc density clipping. For reproducibility, we will report the parameters used in Section 4.2: modulation-network width $12$ with two hidden layers, $\delta=0.55$, Adam learning rate $10^{-2}$ with cosine decay, refresh threshold $\eta_{\mathrm{refresh}}=0.05$, multiscale RBF radii $(0.09,0.17,0.30)$, refresh move limit $0.14$, QP regularization $2\times10^{-7}$, sensitivity-filter width $1.5$ cells, and Armijo constant $10^{-4}$. Accordingly, the revised manuscript will distinguish the exact continuous guarantee from its discretely certified FVM realization, describe the cell-centered construction and feasibility map, and add the representative residual/state-error table and implementation thresholds.
AuthorsOpenReview ↗

Validation of the trade-off between restricted design space and reference refreshing

We thank the reviewer for this suggestion. We added a targeted reduced-symmetry case with seven heat sources and one bottom sink, for which substantial flow-path reorientation is required. The first cheap phase reduces the objective from 434.33434.33 to 254.37254.37 while changing the flux direction by only 0.035∘0.035^\circ on average, consistent with the restriction of a fixed reference family. The near-zero directional change at this endpoint identifies when the restriction becomes consequential: further progress requiring heat-flow-path reorientation cannot be supplied by the fixed reference family alone. After reference refreshes, the mean flux-direction change is 8.90∘8.90^\circ; 32.232.2\\% of active cells rotate by more than 15∘15^\circ. This directly isolates the role of refreshing in changing the heat-flow paths. Moreover, the subsequent optimization, including three accepted refreshes and the intervening cheap phases, reduces the objective from 254.37 to 136.91. This shows that a small number of refreshes can enable substantial further progress beyond the initial fixed reference family.

Table 2. Flow-path change in the targeted reduced-symmetry case.

ComparisonMean angle changeCells above 15∘15^\circ
Initial →\rightarrow first cheap endpoint0.035∘0.035^\circ0.00.0\\%
First cheap endpoint →\rightarrow final8.90∘8.90^\circ32.232.2\\%

We also added the requested matched-budget ablation and supplementary periodic-refresh results. The refresh-only baseline uses the same multiscale RBF representation and constraints, applies full-space-gradient-driven updates, and removes the analytical cheap phase. Table 3 reports the continuous objective at the stated optimization state-solve budget; post-hoc binary audits are excluded from the solve counts. For the completed diagnostic and periodic runs, it additionally reports an independent FVM audit after applying the same hard-binary projection; the two five-solve intermediate checkpoints were not binarized.

Table 3. Matched-budget ablation and supplementary periodic-refresh results.

MethodOptimization solvesContinuous JJBinary JJ
Complete method (five-solve checkpoint)55137.336137.336--
Refresh-only RBF (five-solve checkpoint)55156.738156.738--
Diagnostic-based77136.91136.91141.01141.01
Periodic-601212140.53140.53143.73143.73
Periodic-12055147.17147.17157.87157.87
Periodic-24055149.05149.05217.76217.76
Periodic-30055147.34147.34179.07179.07

At the same budget of five optimization state solves, the complete method achieves a continuous-design objective of 137.336137.336, compared with 156.738156.738 for refresh-only RBF, a reduction of 12.412.4\\%. Because the representation, constraints, and state-solve budget are matched, this comparison isolates the benefit contributed by the analytical cheap phases before hard projection. Binary values are intentionally not reported for these intermediate checkpoints.

For the refresh schedule, the diagnostic-based method achieves J=136.91J=136.91 and a binary-audited objective of 141.01141.01. Periodic-60 uses 12 rather than seven optimization state solves and yields 140.53140.53 and 143.73143.73, respectively; this comparison is not budget-matched. Periodic-120, -240, and -300 give binary-audited objectives of 157.87157.87, 217.76217.76, and 179.07179.07, demonstrating strong sensitivity to the prescribed period. Periodic-120, -240, and -300 have fewer optimization state solves because their subsequent scheduled refresh attempts failed the common acceptance checks; their reported objectives correspond to the last accepted refresh, not to an artificially reduced solve budget. Across the tested schedules, the missing-gradient diagnostic attains the lowest binary FVM-audited objective, but the optimization solve budgets differ; we do not claim a budget-matched advantage over every periodic variant. We will add the targeted flow analysis, matched-budget ablation, periodic comparison, and this qualification to the revised manuscript.

AuthorsOpenReview ↗

Clarification of the practical trade-off between computational cost and design quality

We thank the reviewer for identifying the ambiguity in the original computational comparison and the inconsistent descriptions of the retraining budget. We reran both methods using the same adaptive state-retraining protocol. For every candidate, training is evaluated every 20 Adam steps and terminates once the same held-out accuracy gates are satisfied, with a common maximum of 1,200 steps. The gates cover the PDE residual, heat balance, energy--flux consistency, and the Neumann boundary residual. The previously stated 200-step cap has therefore been removed.

The apparently counterintuitive original comparison of 562562 versus 373373 s arose because each proposed refresh was allowed up to 1,200 Adam steps, whereas each baseline candidate was capped at 200 steps. The proposed timing also included the cheap phases, missing-gradient diagnostics, RBF correction, and state distillation. Thus, training-call counts alone were not a valid proxy for runtime, and we no longer use the original timing comparison as evidence of computational superiority.

Under the revised protocol, there is no failure-based early rejection during state retraining. Both methods are checked every 20 steps and may stop early only by passing all four common gates. A retrained candidate is rejected for state inaccuracy only if it still fails at 1,200 steps; the objective guard and design-feasibility checks are then applied identically. The rejected second cheap endpoint of the proposed run is a cheap-state certification failure before a refresh trial, not an early rejection of a partially trained state network.

We also agree that the original baseline did not demonstrate design stationarity. The revised baseline stops after the prespecified ten accepted design updates, rather than by a stationarity certificate. Its final accepted PINN-objective improvement is 3.58×10−53.58\times10^{-5}, indicating saturation of this trajectory, but we do not claim formal first-order stationarity. We therefore restrict the comparison to achieved design quality under the stated computational protocol and report the stopping reason explicitly.

Table 4. Comparison under the common adaptive retraining protocol.

MethodTraining trials/passedState solvesTime (s)PINN JJFVM JJ
Every-update, original 200-step cap17/617/61818373.5373.5--1.57341.5734
Every-update, common 1,200-step cap15/1015/1016161683.81683.81.29821.29821.54291.5429
Proposed, common 1,200-step cap1/11/1 refresh22251.1251.11.01441.01441.06851.0685

The enlarged budget reduced the baseline rejection rate: 10 of 15 candidates were accepted, compared with 6 of 17 under the original cap. However, its independently audited FVM objective improved only from 1.57341.5734 to 1.54291.5429 (1.94%1.94\%), while its runtime increased from 373.5373.5 to 1683.81683.8 s. Under the common protocol, the proposed method attained an FVM objective of 1.06851.0685 in 251.1251.1 s. Relative to the retrained baseline, this is a 30.75%30.75\% lower objective with 85.09%85.09\% less algorithm time.

The shared reference-network pretraining requires 185.55185.55 s. It is excluded from both algorithm times because the same pretrained reference is used by both methods. If included, the corresponding totals are 436.7436.7 s for the proposed method and 1869.41869.4 s for the retrained baseline, so the conclusion is unchanged. The independent FVM audits are likewise excluded because they are performed only after each design is fixed and never affect either trajectory.

To make the practical trade-off explicit, we re-solved every accepted baseline design and the accepted refresh and termination of the proposed method with the same independent FVM solver. These post-hoc audits do not affect either optimization trajectory. The baseline never reaches the proposed method's audited quality: its best value is J/J0=0.745J/J_0=0.745 at 1683.81683.8 s, whereas the proposed method reaches J/J0=0.516J/J_0=0.516 at 239.6239.6 s and terminates at 251.1251.1 s. Thus, no equal-quality baseline time exists within the observed trajectory; the proposed result dominates in both achieved quality and wall-clock time. All cheap-phase, diagnostic, RBF, distillation, and state-retraining costs are included. We have revised the manuscript accordingly.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.