acceptodds
Under review as a conference paper at ICLR 2027

Latent Search Manifolds: Reusing Control Experience for PDE-Constrained Optimization

11 views

Abstract

High-dimensional PDE-constrained control in repeated-task settings faces two distinct online burdens: evaluating candidate controls can require expensive PDE solves, while optimization must still explore a high-dimensional control space. Learned operators can reduce evaluation cost, but a cheap evaluator does not determine where the optimizer should search and can become unreliable when search leaves well-supported regions. We exploit a different reusable asset: the controls produced by previous optimization tasks. Treating these controls as optimization experience, we organize their shared structure into a continuous, low-dimensional search manifold. A sliced-Wasserstein autoencoder provides the representation needed to reconstruct controls while shaping a latent space that can be searched directly. New tasks are then solved by optimizing their objectives on this manifold rather than predicting controls directly. Representation, optimizer, and evaluator can therefore be chosen separately: the manifold determines where to search, derivative-free optimization determines how to search, and either a numerical solver or a frozen neural operator evaluates candidate controls, with final solutions numerically verified. Across eight heterogeneous PDE-control benchmarks, controlled comparisons show that search geometry alone can change solver-verified outcomes by orders of magnitude: on NS2inlets, the objective decreases from 2.325 in the raw control space to 0.0345 on the learned manifold under the same search budget. The learned search space remains compatible with different derivative-free optimizers and with either exact or surrogate evaluation; when PDE solves are expensive, the latter can further reduce online cost. These results identify reusable control geometry as an independent design axis for repeated PDE optimization, complementary to the choices of optimizer and evaluator.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
PDE-constrained optimization, control manifolds, latent-space optimization, derivative-free optimization, neural operators, scientific machine learning, optimal control

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_NhhYOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The authors proposed to learn low dimensional representation on the spatially and temporal control functions collected from previous optimization task. For a new task, they optimize latent variables and decode them into control functions. They suggests that this low dimensional would make the search of control function in a new task easier as the manifold contains the structures of control. The method is evaluated on eight benchmarks covering static and time-dependent problems, distributed and boundary controls, and regular and irregular meshes. Burgers uses receding-horizon execution. They show that similar order of objective values can be achieved in less time. They validate the methodology by comparing different types of low-dimensional representations, such as DCT and PCA, with a nonlinear autoencoder. For irregular meshes, they using the linear representations like DCT/PCA/GFT followed by autoencoder and show gains in performance.

Critical strengths

The strength of the paper lies in its diverse and extensive empirical studies, which cover uniform-grid and nonuniform-grid problems, time-dependent and static optimization problems, and the use of different optimizers.

Critical weaknesses

The benefit of learning a latent control representation over using a fixed basis is unclear. Distributed controls can already be parameterized using Gaussian or sine/cosine functions, making optimization more manageable. Although the paper includes DCT and PCA baselines, the additional benefit of the autoencoder needs clearer justification. Searching directly over a very high-dimensional control space may make the raw-space baseline unnecessarily difficult. Figures 6 and 7 show substantial differences between the target and optimized states, including for traditional optimization. The authors should clarify whether this mismatch comes from restrictions in the problem formulation or limitations of the optimization method. Table 2 shows that performance depends on the optimization algorithm. It would help to clarify how much improvement comes from the learned representation and how much depends on the optimizer. Is the available data sufficient to learn a useful control manifold? Since historical optimized controls are already available, a comparison with imitation learning would help establish when learning a control search space is preferable to learning a policy.

Questions for authors

Does representing control functions using a fixed basis, such as Gaussian or sine/cosine functions, improve optimization algorithms in general? Would learning a low-dimensional manifold still provide any benefits?

Why were the PDE control benchmark problems presented in PDE Gym and other works on PDE backstepping and differentiable predictive control of PDEs not used? Would your methodology work on one of these example problems?

Given that you have a history of optimization results, would imitation learning work here? Why or why not? Showing imitation-learning results would help establish the usefulness of learning a low-dimensional space.

7 comments
AuthorsOpenReview ↗

Dear Reviewer - Response to NhhY (1/6): Fixed Bases and Learned Search Manifolds

We sincerely thank you for your careful review and thoughtful comments!

We especially appreciate your attention to novelty and what learned search representations offer beyond standard low-dimensional parameterizations. Your questions about Gaussian and sine/cosine bases, optimizers, and direct control prediction point to a fundamental question: What can past optimization experience provide for new tasks beyond generic dimensionality reduction?

Your comments clarified our research motivation and led us to add fixed-basis comparisons and examine physical constraints, representations, and optimizers more carefully. We respond in turn.

1. Are fixed Gaussian or Fourier bases sufficient? Why learn a search manifold?

Thank you for raising this key question. Gaussian, sine/cosine, and other fixed low-dimensional bases are valuable in PDE control: they reduce variables and have clear structure.

Our key distinction is that we are not simply proposing another dimensionality-reduction method. Reduction is a means of searching; we study whether experience from solving related optimization tasks can become a learned, reusable search capability.

Reducing the dimension and learning which controls are worth searching are different problems. Fixed Gaussian or Fourier bases specify the function shapes that can be combined. We instead also learn from high-quality controls found on previous tasks: how control variables co-vary, which combinations repeatedly yield effective solutions, and how to organize this experience into a continuous search space.

A nonlinear decoder can capture correlated changes among valuable control structures beyond prespecified linear basis combinations. But the real test is improved optimization, not just better reconstruction. Following your suggestion, we added direct comparisons.

On nonlinear NS2inlets, we represent 62 free inlet controls with 16-dimensional Gaussian-RBF, Fourier, DCT, PCA, AE, and SWAE coordinates, plus full-space Raw search. Gaussian widths, basis conditioning, and search bounds are calibrated from historical controls, not test targets. All share a frozen DeepONet, Dual Annealing, 200 surrogate evaluations, and top-3 fine-FEM selection.

RepresentationIn-range ↓Boundary ↓Shift ↓Strong shift ↓
Raw2.243932.290442.350212.42324
Gaussian-160.046760.047760.056060.06874
Fourier-160.052670.053280.066100.06830
DCT-160.043210.049320.063320.07534
PCA-160.034760.043930.055180.06549
AE-160.027400.047180.053210.07112
SWAE-160.026030.036390.045430.06840

Mean FEM-verified objectives over three search seeds, for a=0,0.16,0.32,0.48a=0,0.16,0.32,0.48. All use the same surrogate search and top-3 FEM selection. Raw includes the original mean-control penalty.

SWAE gives the lowest mean objective among all 16-dimensional representations for the first three targets. In-range, it reaches 0.02603 versus 0.03476 for the best linear representation, PCA: a 25.1% improvement. Its improvements over PCA remain 17.2% at the training boundary and 17.7% at the first out-of-range target.

At the strongest shift, PCA performs better. This is revealing: historical experience focuses search, but representations cover different control structures, and their advantages can change as new tasks depart from past experience.

We also increased Gaussian and Fourier bases to 32 dimensions and tested prespecified in-range and strongest-shift targets. Larger bases reconstruct historical controls better, yet do not improve final optimization under the same 200-evaluation budget.

Expressing more control shapes does not necessarily make good controls easier to find under a finite budget. This is why learned search manifolds merit study: a good representation should both express valuable controls and make them accessible to optimization.

We additionally compared Gaussian bases on Heat2D and Poisson, where every candidate is scored by the exact numerical PDE solver. These experiments also show search gains from historically learned representations within the tested budgets.

In future work, we hope to combine PCA, Gaussian, and Fourier structures with nonlinear learning, bringing physical priors and historical optimization experience together.

Thank you again for this suggestion. It helped us clarify our contribution: we learn not just lower-dimensional control coordinates, but reusable search structure from past optimization. These results and protocols appear in Appendix E.1–E.2 and Appendix F.3–F.4 (Tables 24–26) of the revised manuscript.

AuthorsOpenReview ↗

Dear Reviewer - Response to NhhY (2/6): Physical Reachability in Figures 6 and 7

2. Why do the optimized states in Figures 6 and 7 still differ visibly from their targets?

Thank you for examining these visualizations so carefully and raising this important question.

In PDE-constrained control, a target state need not be physically attainable under the admissible controls. The optimizer seeks the closest state permitted by the governing PDE, boundary conditions, and control constraints.

Figures 6 and 7 illustrate precisely these limits of physical reachability.

For Heat2D in Figure 6, the desired temperature field varies across both space and time, while the admissible actuation is only a spatially uniform scalar heat source f(t)f(t). At each instant, one source intensity acts throughout the domain; the controller cannot independently heat different spatial locations. Reproducing an arbitrary target field would generally require a source that varies with both position and time, which this benchmark does not allow. Moreover, Figure 6 displays only the snapshot at ts=0.5t_s=0.5, whereas the tracking objective is evaluated over the full space-time domain. The visual difference at a single instant therefore does not represent the entire optimization objective.

For Poisson in Figure 7, the source control is confined to the central circular region Ωc\Omega_c. The state must satisfy u=1u=1 on the outer boundary and u=0u=0 on the boundaries of four interior holes. The cosine target remains positive near those holes, so the desired field is not fully compatible with these prescribed boundary values. The optimizer cannot remove this incompatibility by changing a source restricted to Ωc\Omega_c.

To distinguish physical unattainability from optimizer error more directly, we additionally constructed specialized convex QP reference solutions for the exact target tasks depicted in these two figures, and verified the resulting controls using the original numerical PDE solvers.

For the official Heat2D target, the verified QP objective is 0.036331, closely matching the adjoint reference of approximately 0.0363 displayed in Figure 6. For the Poisson cosine target, the verified QP objective is 3.668484, consistent with the adjoint-FEM result of approximately 3.669 in Figure 7. Both specialized references retain nonzero tracking error even after optimization.

Thus, a substantial part of the visible mismatch arises from the states that the available actuation and boundary conditions can physically reach, rather than simply from insufficient optimization by the learned manifold. This distinction matters when interpreting the figures: a visible target-state gap does not, by itself, establish that the search procedure has failed.

Your observation helped us recognize that these physical constraints needed to be explained more clearly. We have therefore added explicit reachability discussions in Appendix B, immediately before Figures 6 and 7, covering the control restrictions, boundary conditions, the scope of the Heat2D tracking objective, and the solver-verified QP references. These explanations help readers interpret the visualized states in light of the actual PDE-control problems.

We sincerely appreciate your careful attention to these figures, which has helped us make their physical meaning and numerical evidence clearer.

AuthorsOpenReview ↗

Dear Reviewer - Response to NhhY (3/6): Search Representation vs. Optimizer Contributions

###3. How do we separate the contributions of the search representation and the optimizer?

We sincerely thank you for raising this important question. It is precisely why our framework separates the search representation, optimizer, and evaluator into distinct design choices.

An optimization method may find better controls because it uses a more effective search algorithm, or because it searches in a more promising region of the control space from the outset. We want to understand these two contributions separately.

In fact, the original manuscript already included two complementary controlled comparisons.

In Section 4.3 and Figure 3, we fix the optimizer, evaluator, and search budget within each benchmark, changing only the control representation among Raw, DCT, PCA, SWAE, and other applicable structured coordinates. Thus, differences in the resulting objectives isolate the effect of choosing where to search.

In Section 4.4 and Table 3, we instead fix the learned search manifold and compare Dual Annealing, CMA-ES, DANTE, and DANTE-GSA. This examines how different algorithms explore the same search space, separating the optimizer's contribution from the representation itself.

The original Appendix F.2 also repeated the representation comparison using a second optimizer, independently fixed for each benchmark. These results appear in Table 23 of the revised manuscript. For example, with DANTE fixed on NS2inlets, the mean FEM-verified objectives for DCT, PCA, and SWAE are 0.0466, 0.0399, and 0.0332, respectively.

This shows that the benefit of changing the search representation is not confined to one particular optimizer. Even when the search algorithm changes, the structure of the control space exposed to it continues to affect which solutions can be found.

The additional experiments in this revision strengthen this distinction from two further directions. Our new Gaussian/Fourier comparisons on NS2inlets keep the optimizer and evaluation procedure fixed while comparing prescribed and learned search coordinates. Our exact-PDE-only experiments on Heat2D and Poisson remove the surrogate entirely, so every candidate is evaluated by the numerical PDE solver. These comparisons provide further evidence that the search representation has an effect independent of both the choice of optimizer and the surrogate's prediction accuracy.

Search representation and optimizer are therefore not competing explanations. The representation determines which control structures the optimizer can explore; the optimizer determines how that space is explored. Each can influence the final result independently, and the two can also work together.

We have made the connections among these comparisons more explicit in the revised Section 4.1 and Sections 4.3–4.4. The complementary second-optimizer results remain in Appendix F.2 (Table 23), while the new exact-PDE comparison is reported in Section 4.3 (Table 2) and the fixed-basis studies in Appendix F.3–F.4.

We sincerely appreciate your question. It helps us highlight a central methodological point of this work: learning where to search is a research problem in its own right, distinct from designing the algorithm that searches there.

Reviewer_oRnyOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper explores latent search manifolds that reuse prior control experience for PDE-constrained optimization. This paper also identifies a critical waste in many traditional optimization methods: “re-search from scratch for every task. To overcome this waste, this paper trains a reusable sliced-Wasserstein autoencoder that provides a high-quality searchable space. On NS2inlets, the proposed methods improve the objective score in raw control space by an order of magnitude.

Critical strengths
  1. This paper explores the latent search manifolds to optimize search procedures in existing PDE optimization for different tasks, which is interesting and timely.
  2. The idea is well illustrated through the visualization and makes sense.
Critical weaknesses
  1. The paper should improve its presentation. Many acronyms, like BPN and DCT, are never introduced. This makes it hard to focus on the evidence presented.
  2. The main paper does not discuss the autoencoder's training or the rationale for choosing SWAE, which is one of the paper's major contributions.
  3. The experimental setup could be better organized. I would suggest having dedicated sections for benchmarks and baselines.
Questions for authors

See weaknesses.

7 comments
AuthorsOpenReview ↗

Dear Reviewer - Response to oRny (1/3): Acronyms and Method Clarity

We sincerely thank you for your careful reading of our paper and for your concrete, constructive suggestions!

We especially appreciate your attention to clarity. Our central idea is to learn a reusable search space from historically optimized controls. But if readers must repeatedly turn to the appendix to understand why we use SWAE, how the models are trained, or what each experiment actually compares, then we have not communicated that idea clearly enough.

Your comments helped us focus the revision on a clear progression: first, show how historical controls become a new search space; second, explain why SWAE is suited to that role; and finally, make clear how the experiments isolate the effects of search representation, optimizer, and evaluator.

We address your three questions in turn.

1. Acronym definitions and clarity of method presentation

Thank you for pointing out that acronyms such as BPN and DCT were not sufficiently explained when introduced. These terms refer to quite different concepts in PDE optimization and representation learning; presenting them side by side without context makes the paper unnecessarily difficult to follow.

We have clarified the terminology at the relevant places in the revised manuscript. Section 3.1 introduces the discrete cosine transform (DCT), principal component analysis (PCA), and graph Fourier transform (GFT) as structured ways to represent controls, and explains how SWAE learns from historical controls within those coordinates.

Section 4.1 identifies bi-level physics-informed neural networks (BPN) as the optimization reference of Hao et al. (2023), describes PDE-OP as a direct task-to-control prediction approach, and introduces the reference neural operator (RNO) used for irregular meshes. We have also clarified the principal optimizers and evaluators, together with their relevant references. The Figure 2 caption now explicitly explains MSE, FEM, and RNO.

More importantly, we have made the roles of these methods explicit. Where to search is determined by control representations such as DCT, PCA, GFT, and SWAE. How to search is determined by algorithms such as Dual Annealing, CMA-ES, and DANTE. How to evaluate candidate controls is determined by the numerical PDE solver or a frozen learned operator.

These three choices are now organized together in Section 4.1 and examined separately in Sections 4.3–4.5, so readers can understand both why each method appears in the experiments and how it differs from the others.

We have also strengthened the explanation of SWAE in the Introduction and Section 3.1. Reconstruction preserves the structure of historical controls, while sliced-Wasserstein regularization organizes the latent distribution for search. The online search bounds are then determined separately from historical latent encodings. Appendix E provides the detailed representation-training and evaluation settings.

Our goal is for readers to understand the logic of the framework from the main text, with the appendix supporting reproducibility rather than supplying missing conceptual connections.

Thank you again for helping us communicate the idea more clearly. The search space is itself an optimization design choice, distinct from the algorithm that explores it and the evaluator that scores candidate controls. Making this distinction easy to follow is central to the contribution we hope the paper conveys.

AuthorsOpenReview ↗

Dear Reviewer - Response to oRny (2/3): Why SWAE and How We Learn the Search Space

2. Why SWAE, and how is the search space learned from historical controls?

We sincerely appreciate your request for a clearer explanation of why we chose SWAE. This is one of the most important parts of understanding our method, and we wanted to make it more intuitive in the main text.

Historically optimized controls are more than a collection of answers. When good controls recur across related tasks, they reveal structures shared by high-quality solutions. Our goal is to organize those structures into a new search space in which new objectives can be optimized directly.

A standard autoencoder (AE) already provides an essential capability: it compresses high-dimensional controls into continuous coordinates and reconstructs physical controls through a deterministic decoder. This allows new tasks to be optimized in a lower-dimensional space rather than adjusting every control variable from scratch.

But online optimization does not visit only the latent codes seen during training. It continually proposes new coordinates between and around historical embeddings. Beyond reconstructing existing controls, we therefore want the aggregate latent distribution to be better organized for continued search.

This is why we use a sliced-Wasserstein autoencoder (SWAE). SWAE combines control reconstruction with sliced-Wasserstein distribution regularization: reconstruction preserves the physical structure of historical controls, while regularization organizes the aggregate latent-code distribution toward a simple prior. After training, empirical quantiles of the historical encodings define the bounded online search region.

Historical controls determine where to search; distribution regularization helps organize the search coordinates; the new task's objective determines which control is ultimately selected.

The representation is learned once offline. It can then be used with different derivative-free optimizers and either the numerical PDE solver or a learned surrogate, without retraining the control representation.

We have made this process explicit in revised Section 3.1, from historical control experience and encoding/reconstruction to sliced-Wasserstein regularization and search bounds derived from training codes. Appendix E gives the full network settings, data splits, and model-selection details.

We also added a matched AE–SWAE experiment to test directly what distribution regularization contributes.

For nonlinear NS2inlets, we collect 450 control snapshots from optimized trajectories for 30 historical auxiliary objectives, representing 62 free inlet-control values with 16 latent coordinates. AE and SWAE use the same training controls, network architecture, paired initial weights, training duration, and search budget. The key difference is SWAE's additional sliced-Wasserstein regularization.

We freeze both representations and evaluate them on four progressively changed targets:

NS2inlets targetRaw ↓AE ↓SWAE ↓
In-range reference2.243930.027400.02603
Training boundary2.290440.047180.03639
Noticeable shift2.350210.053210.04543
Strong shift2.423240.071120.06840

Mean FEM-verified objectives over three paired training and search runs. Each search uses 200 surrogate evaluations, then selects the best of the top three candidates using FEM. Raw retains the original mean-control penalty.

SWAE obtains a lower mean numerical objective than AE at all four target levels. Its improvement over AE is approximately 22.9% at the training boundary and 14.6% at the first target beyond the historical training range.

These results matter to us because AE already learns a continuous, low-dimensional control representation. SWAE's distribution regularization further improves search on this nonlinear flow-control problem. This provides direct experimental support for our choice of SWAE and clarifies our central point: we care not only about how many dimensions a control is compressed into, but whether the resulting space actually helps an optimizer discover new solutions.

We have highlighted this design rationale in the revised main text, while retaining the complete matched-training protocol in Appendix E and the target-wise results in Appendix F.4 (Table 25).

AuthorsOpenReview ↗

Dear Reviewer - Response to oRny (3/3): Benchmark, Baseline, and Experimental Comparisons

Benchmark, Baseline, and Experimental Comparison Structure

Thank you very much for pointing out how the experimental setup could be better organized. We use eight PDE-control benchmarks to examine the applicability of our method across different dynamics, control settings, and geometries. However, these benchmarks involve baselines from different sources and different numerical evaluation procedures. Without first establishing an overall structure, it is indeed difficult for readers to understand what each comparison is intended to show.

We have therefore reorganized Section 4.1 to introduce the eight benchmarks by their origins and physical characteristics, before explaining the baselines, search representations, optimizers, and evaluation methods.

Specifically, Voltage and Burgers follow the control problems associated with PDE-OP, covering static spatial control and nonlinear spatiotemporal trajectory control.

Heat2D, Poisson, and NS2inlets follow the PDE-control formulations studied in the BPN work, covering heat conduction, an elliptic PDE on a perforated domain, and nonlinear Navier–Stokes boundary control.

Irregular Heat, Darcy, and Advection form our newly generated FEM control suite, drawing on the operator-optimization setting of Opt-RNO. These tasks allow us to study search representations across different PDEs and irregular geometries.

We then explain the roles of classical numerical optimization methods, existing learning-based baselines, structured low-dimensional representations, derivative-free search algorithms, and exact or surrogate evaluators. This organization also makes clear that our experiments investigate three related but separately testable questions.

First, what changes when we change where the optimizer searches? Figure 3 and Section 4.3 hold the optimizer, evaluator, and search budget fixed while comparing representations such as Raw, DCT, PCA, and SWAE. This isolates the effect of the search space itself.

Second, what changes when different optimizers explore the same search space? Section 4.4 and Table 3 fix the learned representation and compare Dual Annealing, CMA-ES, DANTE, and DANTE-GSA. The results distinguish the effect of the optimizer from the effect of the coordinates in which it searches.

Finally, what changes when candidate controls are evaluated by a numerical PDE solver rather than a learned surrogate? Section 4.5 and Table 4 examine this question with the representation and optimizer fixed, comparing solution quality and computational cost while numerically verifying the final controls.

We have also clarified the distinction between Table 1, which presents solution quality and online-time comparisons across methods, and the subsequent controlled experiments, which use fixed evaluation budgets, fixed representations, or matched computational time as appropriate.

The revised Section 4.1 now explains the conditions under which different evaluators are used, how final controls are numerically verified, and where the historical-control and operator-training data come from. In particular, the optimized-control banks and operator-training data used by our models are generated with numerical PDE solvers rather than taken directly from the released benchmark datasets.

We believe these changes make the purpose of each experiment much easier to follow. Readers can now understand what is being held fixed, what is being changed, and what each comparison tells us, without first needing to be familiar with every algorithm or benchmark family.

Thank you again for helping us make the experimental evidence more accessible and its connection to the central idea of the paper more explicit.

Reviewer_8GanOpenReview ↗
Rating
3
Confidence
3(+1)
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a framework for repeated PDE-constrained optimization that learns a low-dimensional search manifold from previously optimized controls using a sliced-Wasserstein autoencoder. For new tasks, derivative-free optimization is performed in the learned latent space, with candidates evaluated by either a numerical PDE solver or a neural-operator surrogate. The framework is evaluated on 8 PDE-control benchmarks covering different dynamics, control settings, and geometries. The results show that learned search representations can substantially improve optimization efficiency and solution quality under limited evaluation budgets, while maintaining numerical verification of the final solutions.

Critical strengths
  1. The paper focuses on reusing historical control solutions to improve the search space itself, which is an important but relatively underexplored aspect of repeated PDE optimization. Separating the search representation from the optimizer and evaluator provides a flexible framework.

  2. The experiments cover 8 PDE-control benchmarks, including nonlinear dynamics, boundary and distributed controls, trajectory optimization, and irregular geometries. The comparisons across Raw, DCT, PCA, and SWAE representations, as well as different optimizers and evaluators, help isolate the contribution of search-space design. The additional ablations and sensitivity analyses strengthen the empirical evidence.

  3. The results show substantial improvements over raw-space search under matched evaluation budgets. In several benchmarks, latent optimization achieves solution quality close to classical numerical methods at significantly lower reported online cost. The framework also supports multiple derivative-free optimizers and both exact and surrogate evaluation, making the proposed approach practically efficient and flexible.

Critical weaknesses
  1. Several benchmarks appear to admit more specialized optimization formulations. For instance, Voltage, Heat2D, and Poisson can be formulated as convex quadratic programs, while the irregular Heat and Darcy problems appear to have linear objectives in the controls under the stated constraints. Comparisons with specialized solvers (e.g., OSQP or HiGHS), or more efficient adjoint-based optimization implementations, would help strengthen the evaluation. Besides, a clearer explanation of the substantial wall-clock differences between latent search and classical optimization would help establish the practical efficiency gains. Could the authors clarify the classical optimization implementations and provide comparisons with specialized convex solvers under matched computational budgets?

  2. In the surrogate-based experiments, learned representations naturally constrain optimization to regions better covered by surrogate training data. Thus, improvements over raw-space search may partly reflect reduced surrogate exploitation rather than superior search geometry. More controlled comparisons, particularly with exact numerical evaluation, would help disentangle these effects. Could the authors clarify how much of the improvement comes from search-space geometry itself, rather than differences in surrogate accuracy across representations?

  3. Optimizing within a learned manifold may exclude better solutions outside its range. Although the paper examines distribution and geometry shifts, the robustness of SWAE under such shifts remains insufficiently explored. Could the authors further evaluate SWAE under distribution shifts and investigate whether full-space refinement can overcome this limitation?

Questions for authors

See weakness

10 comments
AuthorsOpenReview ↗

Dear Reviewer - Response to 8Gan (1/5): Convex Solver Comparison

We sincerely thank you for your generous and encouraging assessment of our work and for your thoughtful engagement with its central idea! We especially appreciate your recognition of the value of reusing optimization experience in PDE-constrained control. Your suggestions have helped us strengthen both the experimental evidence and the motivation of the paper.

1. Specialized convex solvers and computational efficiency

Thank you for this important suggestion. You are right that some PDE-control problems have exploitable convex structure, and that specialized solvers provide valuable numerical references for evaluating our approach.

Following your suggestion, we formulated Voltage, Heat2D, and Poisson as box-constrained convex quadratic programs (QP), solved them using OSQP, and verified the controls with the original numerical PDE solvers. We also added linear-programming (LP) experiments on irregular Heat and Darcy, solving six additional FEM instances for each PDE with HiGHS and verifying their results numerically.

These experiments allow us to ask an important practical question: As control problems move from relatively simple structures toward higher dimensions, more demanding PDEs, and complex geometries, how much of the quality and efficiency advantage of specialized convex solvers remains?

BenchmarkQP objective ↓SWAE objective ↓QP online timeSWAE online time
Voltage0.0008600.006140 (surrogate)0.031 s0.30 s†
Heat2D0.0356350.036662 (exact)1.034 s1.848 s
Poisson4.7562664.847578 (exact)24.108 s0.477 s

SWAE values are from the original Table 1. † Voltage reports surrogate-search time; five numerical confirmations are additional.

On Voltage, the QP advantage is striking. Its relatively simple control structure allows a specialized convex solver to exploit the known mathematical relations directly, achieving a more accurate control in only 31 milliseconds. For this kind of problem, a dedicated QP solver is clearly an excellent choice.

On Heat2D, the advantage becomes considerably smaller. The SWAE objective is only about 2.9% above the QP reference, and their online times are already of the same order. QP still leads, but the large quality gap seen on Voltage has nearly disappeared.

On Poisson, with 509 control degrees of freedom and a perforated computational domain, the efficiency comparison changes again. QP still reaches the lower objective, but SWAE comes within approximately 1.9% of its reference. The QP online time is 24.108 seconds, compared with only 0.477 seconds for the original exact-PDE SWAE search. Here, learned search reaches nearly the same quality at substantially lower online cost.

Together, these examples trace a clear progression: QP strongly leads in both accuracy and speed on Voltage; its quality advantage narrows and the online costs approach the same scale on Heat2D; on Poisson, QP retains the better objective but SWAE takes the efficiency advantage.

An additional point is important for interpreting this comparison fairly. The SWAE representations reported here were trained on historical controls generated by our original numerical optimization pipelines, including adjoint-based methods, rather than on historical controls optimized by the specialized QP solver. Thus, while QP provides a stronger solution-quality reference on these problems, SWAE has not yet been given the opportunity to learn from that stronger source of optimization experience.

This distinction matters because the quality of historical controls can directly influence the search structure that SWAE learns. Regenerating the historical control bank using QP solutions and retraining the manifold could provide a stronger learned search space and potentially further improve performance on new tasks. We have not conducted this additional retraining experiment, so we do not claim a specific improvement. However, we respectfully ask that the present results be interpreted with this distinction in mind: they demonstrate the strength of specialized QP optimization, but do not establish an upper bound on what reusable, experience-based search can achieve.

AuthorsOpenReview ↗

Dear Reviewer - Response to 8Gan (2/5): Applicability and Motivation

1. Specialized convex solvers and computational efficiency (continued)

These are only the three benchmark problems for which we can construct the corresponding convex QP formulations.

For more complex PDE-control problems, the first difficulty with specialized QP/LP methods is often not their speed, but whether the original problem admits an appropriate convex formulation at all.

Among our eight PDE-control benchmarks, we constructed QP references for three problems and additional LP instances for two irregular PDE families. Burgers, NS2inlets, and irregular Advection have more complicated nonlinear control-to-state relations and cannot be handled directly by these same convex formulations. Applying comparable specialized methods would require further linearization, approximation, or a different optimization design, whose applicability and performance would need to be established separately.

This distinction matters in practice. We are particularly interested in problems with expensive PDE evaluations, high-dimensional controls, nonlinear dynamics, and repeated optimization as objectives change. In such settings, convenient convex structure may not exist, and a specialized formulation developed for one PDE-control system may be difficult to transfer to another with different dynamics or actuation.

Historical optimization experience offers another kind of reusable structure.

Our framework learns a search space from high-quality controls found on previously solved tasks within the same family. For each new objective, we optimize the objective within this learned space. This principle does not require a linear control-to-state map or a convex quadratic objective, so it can be applied across different PDE dynamics, control parameterizations, and numerical evaluators.

This is the central challenge our paper seeks to address: when a practical control problem is too complex for a convenient specialized convex formulation, or when that formulation's computational advantage becomes less pronounced, experience from solving related tasks can still tell us where to search next.

Your suggestion has allowed us to make this motivation much more direct. Learned search is not simply another solver for a single simple convex problem: it explores a reusable optimization capability that can support a broader range of challenging PDE-control tasks.

We have now incorporated the QP/LP formulations and numerical verification, online computational costs, and implementation details into the revised manuscript (Appendix G, Tables 34 and 35). These results provide a stronger quantitative foundation for the comparison. We sincerely appreciate your suggestion and the opportunity to clarify the broader role of reusable control experience.

AuthorsOpenReview ↗

Dear Reviewer - Response to 8Gan (3/5): Search Space Benefits Beyond Surrogate Accuracy

2. Does the advantage of the search space come from surrogate prediction accuracy?

We are especially grateful for this insightful and perceptive question!

You have identified a possibility that bears directly on the paper's central contribution: the learned manifold may keep the optimizer in regions where the surrogate predicts more reliably. Does the observed improvement therefore come from the search space itself, or mainly from reduced surrogate prediction error?

This distinction matters because it determines whether historical optimization experience has taught us a structure that genuinely helps search on new tasks.

To answer it directly, we added experiments with no surrogate at all: every candidate control is evaluated by the numerical PDE solver.

In the original Figure 3, the Heat2D and Poisson representation comparisons used surrogate-guided search followed by numerical verification of the final candidates. In the new experiments, we fix the target tasks, search seeds, Dual Annealing optimizer, and evaluation budget, changing only the search space. Heat2D uses 500 numerical PDE evaluations; Poisson uses 200 FEM evaluations.

Search spaceHeat2D ↓Poisson ↓
Raw0.064147.44796
DCT0.0402411.22751
PCA0.036974.81579
SWAE0.036564.85192

Mean objectives computed by the numerical PDE solvers. Each benchmark includes three targets, with three search seeds per target.

Even after removing the surrogate entirely, the learned search space still substantially improves the outcomes of true PDE optimization.

Under the same evaluation budgets, SWAE reduces the mean objective relative to Raw by approximately 43% on Heat2D and 35% on Poisson. More importantly, SWAE outperforms both Raw and DCT in all 18 target-and-seed paired comparisons across the two benchmarks.

Every candidate objective in these experiments comes directly from the numerical PDE solver. The improvement therefore does not depend on the surrogate happening to predict more accurately in some region; it comes directly from the choice of search space.

This further establishes the paper's most important research claim: the search space is an independent optimization design decision. It determines which control structures receive the optimizer's limited computational budget and, consequently, which solutions the optimizer can reach.

These results also reinforce our conviction that historical optimization experience deserves to be studied as a resource in its own right. Previously optimized controls are not merely a collection of high-quality answers. They also contain information about how those answers are distributed in control space and which structures can be reused across tasks.

This is precisely the direction we hope the paper will encourage: learning effective search regions from historical control distributions, understanding the search process through the geometric organization of latent space, and developing search representations better suited to different PDE tasks.

At the same time, your question about surrogate reliability reveals another value of the learned manifold.

As reported in Section 4.5 of the original paper, in Voltage's Top-5 numerical reranking experiment, 77.8% of Raw-search runs changed their initially selected candidate after numerical verification, compared with only 8.3% of SWAE runs.

This shows that, in that experiment, the surrogate's candidate ranking on the learned manifold agrees more closely with numerical evaluation. In other words, search geometry can independently improve optimization with the true PDE, while also making surrogate-guided search more reliable.

We sincerely thank you for helping us distinguish these two effects more clearly. The revised manuscript now reports the exact-PDE experiments in Section 4.3 (Table 2) and Appendix F, alongside the representation comparison in Figure 3 and the numerical reranking evidence in Section 4.5.

Comments on the paper

3 comments
AuthorsOpenReview ↗

Dear Area Chair and Reviewers - Substantial Revisions and Our Sincere Thanks

Dear Area Chair and Reviewers,

We sincerely thank all three reviewers for the time, care, and intellectual rigor devoted to our work, and the Area Chair for considering our revised submission. We have treated your comments not merely as questions to answer, but as an opportunity to examine the central idea of our paper more deeply. We have now substantially revised the manuscript and provided detailed responses in the individual review threads. These revisions go far beyond presentation: they include new experiments, independent numerical references, and more direct tests of what learning a reusable search space truly contributes.

In particular, we added specialized QP references for Voltage, Heat2D, and Poisson, together with 12 newly constructed LP instances on irregular Heat and Darcy, verified using numerical PDE solvers. We introduced exact-PDE-only search experiments, removing the surrogate entirely and showing that SWAE reduces the mean objective relative to Raw search by approximately 43% on Heat2D and 35% on Poisson. We conducted calibrated Gaussian/Fourier fixed-basis comparisons, matched AE–SWAE studies, and progressively shifted-target experiments on nonlinear NS2inlets. We also examined larger fixed bases, the coverage limits of learned manifolds, and further optimization in the original control space. Beyond these experiments, we clarified the physical reachability constraints behind Figures 6 and 7, reorganized the presentation of our eight benchmark families and baselines, and expanded the explanations of training, numerical verification, and computational costs. Together, these additions more clearly distinguish where to search, how to search, and how to evaluate, while revealing both the value and the limits of historical optimization experience.

We would also like to explain why the revised manuscript and detailed responses are arriving together only at this stage. Many of the questions required substantially more than additional writing: we needed to reconstruct mathematical optimization formulations, build new FEM tasks, align experimental protocols, repeat numerical computations, and carefully verify the results before incorporating them into the paper. Over the past several days, we have worked through long days and late nights, often with very little rest, to complete these revisions as quickly and carefully as possible. We are sharing the completed work at the earliest point at which we could responsibly verify it. We sincerely hope you will understand the timing in this spirit. Our priority has been to provide substantive evidence and accurate answers, rather than hurried promises.

We also wish to share something more personal. As researchers, we have been genuinely moved by the depth and seriousness with which you engaged with our work. There is something deeply meaningful about seeing other researchers devote such careful thought to an idea we have worked so hard to develop. Some of your questions were difficult, but precisely because they were difficult, they pushed us to revisit our assumptions, conduct stronger experiments, and think more clearly about what our research is really trying to contribute. We have felt both challenged and encouraged throughout this process. That kind of scientific engagement means more to us than we can easily express, and we are sincerely grateful for it.

Finally, we want to make a clear commitment to the research community: if this paper is accepted, we will publicly release the code for all benchmark implementations and methods developed for this study, together with detailed dataset-construction procedures and scripts, numerical PDE solver settings, control parameterizations, training and search configurations, and numerical verification protocols. We hope these resources will make our results fully reproducible and help other researchers build on, scrutinize, and extend this work.

Thank you once again for your time, your careful questions, and your constructive criticism. We sincerely hope that the new evidence and substantial revisions address your concerns, and we would be deeply grateful if you would consider these improvements when reassessing the paper's scientific contribution and overall quality. Your engagement has genuinely helped us make this a stronger piece of research.

AuthorsOpenReview ↗

Final Reminder to Reviewers

Dear Reviewers,

With only four hours remaining, we kindly urge the reviewers who have not yet responded to revisit our rebuttal and revised manuscript.

We have carefully addressed the concerns with detailed clarifications and additional experimental evidence, and would greatly appreciate your final feedback and updated assessments before the discussion closes.

Thank you sincerely for your time and consideration.

Best regards, The Authors

AuthorsOpenReview ↗

Final Urgent Request for Reassessment

Dear Reviewers,

With only one hour remaining, we sincerely ask for your final consideration of our revised manuscript and rebuttal.

We have made every effort to address the key concerns through rigorous additional experiments and targeted revisions.

If our revisions have substantively resolved your concerns, we respectfully hope this progress can be reflected in an updated rating and recommendation before the deadline.

We deeply appreciate your time, thoughtful feedback, and commitment to a fair, evidence-based evaluation.

Best regards, The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.