acceptodds
Under review as a conference paper at ICLR 2027

Learning Fast Approximate Solvers for PDE-Constrained Optimization

Abstract

Partial differential equation (PDE)-constrained optimization arises in many scientific and engineering domains, such as energy systems, fluid dynamics, and material design. In these problems, the decision variables are tightly coupled with the PDE state variables, and the feasible set is implicitly defined by the governing PDE constraints. This coupling makes these problems computationally demanding, as conventional numerical methods discretize the governing PDE into high-dimensional algebraic systems that must be repeatedly solved throughout the optimization process. To address these challenges, this paper introduces a novel learning-based framework to approximate solutions to PDE-constrained optimization tasks in near real time. The proposed method integrates a dynamic predictor with an optimization surrogate. The dynamic predictor, a time-discrete Neural Operator Lu_2021, efficiently approximates system trajectories governed by PDE dynamics, while the optimization surrogate leverages proxy optimizer techniques kotary2021end to approximate the associated optimal decisions. This dual-network design enables real-time approximation of optimal strategies while explicitly capturing the coupling between decisions and PDE dynamics. We validate the proposed approach on benchmark PDE-constrained optimization tasks including Burgers' equation, heat equation, and voltage regulation, and demonstrate that it achieves solution quality comparable to state-of-the-art control-based algorithms such as the Direct Method and Model Predictive Control, while achieving computational speedups of up to four orders of magnitude.

Primary area:
neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)
Keywords:
Optimal Control, Partial Differential Equations, Neural Surrogates

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_2FZiOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

Summary

The paper proposes PDE-OP (Optimization Proxy), a learning-based framework that solves spatiotemporal PDE-constrained optimal control problems in real time. It replaces slow, iterative test-time optimization (such as MPC or Adjoint Methods) with an amortized single-pass neural rollout. The system consists of two neural networks trained in two stages:

  1. Dynamic Predictor (Yθ\mathcal{Y}_\theta): A discrete-time Neural Operator (inspired by DeepONet) trained offline on arbitrary forward PDE simulations to model system state transitions.

  2. Surrogate Controller (Uω\mathcal{U}_\omega): A policy network trained on-the-fly via self-supervised primal-dual learning to output spatial basis coefficients for control actions. It unrolls predictions through the frozen Yθ\mathcal{Y}_\theta while dynamically penalizing physical and actuator constraint violations via Lagrange multipliers.


Main Contributions

  • Dual-Network Architecture: Integrates a learned discrete-time neural operator with a proxy controller for real-time control rollouts.

  • Discrete-Time Neural Operator: Models direct state-transition maps (yk,ck)↦yk+1(y_k, c_k) \mapsto y_{k+1} without requiring numerical time-integration schemes.

  • Label-Free Primal-Dual Training: Uses self-supervised Lagrangian relaxation to enforce dynamics, boundary conditions, and actuator bounds during training without needing precomputed optimal control.

  • Significant Computational Acceleration: Achieves decision quality comparable to classical solvers while offering orders-of-magnitude faster inference.


Validation

The framework was evaluated on three optimal control benchmarks across open-loop and closed-loop settings:

  • 1D Voltage Equation (Linear Reaction-Diffusion): Evaluated under open-loop control. Achieved comparable accuracy to Direct and Adjoint methods while running in 0.035s (~1.1×103×1.1 \times 10^3 \times faster than the Direct Method).

  • 2D Heat Equation (Linear Reaction-Diffusion): Evaluated under closed-loop control on 2D spatial domains. Matched LMPC and Adjoint tracking accuracy while running in 0.031s (~200×200\times to 2200×2200\times speedup).

  • 1D Viscous Burgers' Equation (Nonlinear Advection-Diffusion): Evaluated under closed-loop control. Produced state trajectories comparable to NMPC and Adjoint methods while running in 0.062s (~1.42×104×1.42 \times 10^4 \times faster than NMPC and ~1.66×103×1.66 \times 10^3 \times faster than Adjoint).

Critical strengths

1. Remarkable Computational Acceleration for Real-Time Execution

The dual-network architecture replaces iterative numerical optimization at test time with an amortized, single-pass neural rollout. Across diverse benchmark tasks—including 1D reaction-diffusion voltage dynamics, 2D heat equations, and 1D viscous Burgers' equations—PDE-OP achieves solution quality comparable to state-of-the-art classical baselines (e.g., Adjoint-Sensitivity Methods and Nonlinear MPC) while providing computational speedups of up to three to four orders of magnitude (e.g., running in 0.062s versus 878.67s for NMPC on 1D Burgers' equation).


2. Self-Supervised Primal-Dual Training Without Optimal Decision Labels

To train the surrogate controller, the framework employs a self-supervised primal-dual formulation based on Lagrangian duality. It updates Lagrange multipliers (λ\lambda) iteratively during training to dynamically penalize violations of initial/boundary conditions, system dynamics, and control bounds. Crucially, this allows the controller to be optimized end-to-end through differentiable rollouts of the neural operator Yθ\mathcal{Y}_{\theta} without requiring precalculated optimal control.

Critical weaknesses

Weakness 1: Inflexibility to Changes in Objective Functions (JJ) and Hyperparameters

The surrogate controller Uω\mathcal{U}_{\omega} is trained via self-supervised primal-dual optimization using a fixed, hardcoded objective function J(u^,y^)\mathcal{J}(\hat{u}, \hat{y}).

Equation (4) defines Uω\mathcal{U}_{\omega} as a mapping

(y(tk,x),ytarget(x))↦c(tk)(y(t_k, x), y_{\text{target}}(x)) \mapsto c(t_k).

It does not take the weights/parameters of the cost functional J\mathcal{J} (such as running cost weight λ\lambda, control effort penalty γ\gamma, or terminal weight) as inputs.


Weakness 2: Two-Stage Training vs. Misleading "End-to-End" Framing

In Section 4 (and abstract/contributions), the paper presents the method as an "end-to-end" learning-based framework / differentiable surrogate. However, the actual pipeline is strictly two-stage:

  1. Yθ\mathcal{Y}_{\theta} is trained offline on synthetic ground-truth numerical PDE trajectories using supervised learning (Eq. 7) and then frozen.

  2. Uω\mathcal{U}_{\omega} is trained on-the-fly through differentiable rollouts of the frozen θ\theta (Eq. 8, 10).


Weakness 3: Overbroad Title and Mismatched Scope vs. Missing Related Work

The paper is titled "Learning Fast Approximate Solvers for PDE-Constrained Optimization". However, the proposed framework, formulation, and benchmark experiments focus strictly on spatiotemporal PDE-constrained optimal control problems (e.g., tracking target profiles ytarget(x)y_{\text{target}}(x) via time-stepping control trajectories u(t,x)u(t,x)). Optimal control is only a specific subset of the broader field of PDE-constrained optimization, which encompasses parameter identification, shape optimization, and topology design.

If the authors insist on positioning the paper under the broad umbrella of general PDE-constrained optimization, the literature review is substantially inadequate. Crucial foundational and recent work on learning-based PDE-constrained optimization—such as self-supervised operator learning for parametric optimization (Wang et al., 2021) and exact gradient-accelerated operator surrogates for PDE optimization (Cheng et al., ICML 2025)—is entirely missing.

[1] Wang, S., Bhouri, M. A., and Perdikaris, P. 2021. Fast pdeconstrained optimization via self-supervised operator learning.

[2] Ze Cheng, Zhuoyu Li, Wang Xiaoqiang, Jianing Huang, Zhizhou Zhang, Zhongkai Hao, and Hang Su. 2025. Accelerating PDE-Constrained Optimization by the Derivative of Neural Operators. In International Conference on Machine Learning. PMLR, 10090–10105.

Questions for authors

N.A.

6 comments
AuthorsOpenReview ↗

Authors' response (1/2)

We thank the reviewer for their detailed assessment of our work, constructive feedback, and recognition of the computational efficiency and self-supervised learning contributions of PDE-OP. We address each of the raised concerns below and clarify the scope and methodological contributions of our framework.

Weaknesses.

Weakness 1: Inflexibility to Changes in Objective Functions and Hyperparameters. The surrogate controller Uω\mathcal{U}{\omega} is trained via self-supervised primal-dual optimization using a fixed, hardcoded objective function. Equation (4) defines Uω\mathcal{U}_{\omega} as a mapping from the state and desired target to control coefficients. It does not take the weights/parameters of the cost functional (such as running cost weight, control effort penalty, or terminal weight) as inputs.

Response. We thank the reviewer for this observation. PDE-OP is designed to learn an approximate solver for a given PDE-constrained optimal control problem, where the objective weights are fixed while initial conditions and target profiles vary across instances.

Nevertheless, the framework can naturally accommodate families of control problems with varying objectives by including the corresponding objective parameters (e.g., running-cost and control-effort weights) as additional inputs to the controller, similarly to how it is conditioned on the target profile and system state. The controller could then be trained over a distribution of objective parameters using the same self-supervised primal–dual formulation, without requiring optimal-control labels. This represents an interesting direction for future works.

Weakness 2: Two-Stage Training vs. Misleading “End-to-End” Framing. In Section 4 (and abstract/contributions), the paper presents the method as an “end-to-end” learning-based framework / differentiable surrogate. However, the actual pipeline is strictly two-stage: Yθ\mathcal{Y}{\theta} is trained offline on synthetic ground-truth numerical PDE trajectories using supervised learning (Eq. 7) and then frozen. Uω\mathcal{U}{\omega} is trained on-the-fly through differentiable rollouts of the frozen Yθ\mathcal{Y}{\theta} (Eq. 8, 10).

Response. We thank the reviewer for this observation. Our use of the term end-to-end was intended to emphasize that both the dynamics surrogate and the controller are differentiable, allowing gradients of the control objective to propagate through the learned PDE dynamics during controller training.

As correctly noted by the reviewer, PDE-OP adopts a two-stage training procedure: the neural operator Yθ\mathcal{Y}{\theta} is first trained on numerical PDE trajectories and subsequently frozen, while the controller Uω\mathcal{U}{\omega} is trained through differentiable rollouts of the pretrained operator. This separation allows the dynamics model to be learned independently and subsequently used for self-supervised controller optimization without requiring optimal-control labels.

We appreciate the opportunity to clarify this distinction. The revised manuscript now explicitly describes PDE-OP as a two-stage training framework with end-to-end differentiable controller optimization, rather than implying joint training of both networks.

Weakness 3: Overbroad Title and Mismatched Scope vs. Missing Related Work. The paper is titled “Learning Fast Approximate Solvers for PDE-Constrained Optimization”. However, the proposed framework, formulation, and benchmark experiments focus strictly on spatiotemporal PDE-constrained optimal control problems (e.g., tracking target profiles via time-stepping control trajectories). Optimal control is only a specific subset of the broader field of PDE-constrained optimization, which encompasses parameter identification, shape optimization, and topology design. If the authors insist on positioning the paper under the broad umbrella of general PDE-constrained optimization, the literature review is substantially inadequate. Crucial foundational and recent work on learning-based PDE-constrained optimization—such as self-supervised operator learning for parametric optimization (Wang et al., 2021) and exact gradient-accelerated operator surrogates for PDE optimization (Cheng et al., ICML 2025)—is entirely missing. [1] Wang, S., Bhouri, M. A., and Perdikaris, P. 2021. Fast PDE-constrained optimization via self-supervised operator learning. [2] Ze Cheng, Zhuoyu Li, Wang Xiaoqiang, Jianing Huang, Zhizhou Zhang, Zhongkai Hao, and Hang Su. 2025. Accelerating PDE-Constrained Optimization by the Derivative of Neural Operators. In International Conference on Machine Learning. PMLR, 10090–10105.

AuthorsOpenReview ↗

Authors' response (2/2)

Response. We thank the reviewer for this observation and agree that the scope of our work should be stated more precisely. PDE-OP specifically addresses PDE-constrained optimal control, which represents an important class of PDE-constrained optimization problems. We therefore propose the more precise title Learning Fast Approximate Solvers for PDE-Constrained Optimal Control and have clarified this scope throughout the revised manuscript.

We also thank the reviewer for highlighting the works of Wang et al. (2021) and Cheng et al. (2025), which we have incorporated into the Related Work discussion. Both studies demonstrate the potential of neural operators for accelerating PDE-constrained optimization. However, while these approaches employ learned operators to accelerate numerical optimization, PDE-OP learns a surrogate controller that directly predicts control decisions for unseen problem instances, avoiding instance-specific optimization at inference time.

We emphasize that our contribution is not the use of neural operators for PDE-constrained optimization itself, but their integration with self-supervised primal–dual controller training to enable fast, amortized solutions to PDE-constrained optimal control problems.

Proposed manuscript addition

Revised title.

Learning Fast Approximate Solvers for PDE-Constrained Optimal Control.

Related Work clarification.

Wang et al. (2021) leverage physics-informed neural operators as differentiable surrogates for PDE-constrained optimization, while Cheng et al. (2025) improve neural-operator derivative accuracy and combine learned surrogates with numerical optimization. In contrast, PDE-OP focuses on amortized PDE-constrained optimal control, learning a controller through differentiable dynamics predictions and self-supervised primal–dual training. Once trained, the controller directly predicts control actions for unseen instances without requiring numerical optimization at inference time.

AuthorsOpenReview ↗

Authors' comment

Thank you again for your review. We wanted to follow up on our clarification that PDE-OP uses two-stage training, addresses PDE-constrained optimal control, and learns a controller for fixed objective weights in the current setting. We revised the title and related-work discussion accordingly and explained how PDE-OP differs from approaches that use learned operators within instance-specific numerical optimization. We also identified conditioning on objective parameters as a possible extension. Please let us know if any part of these scope and training clarifications needs further explanation. If these responses address your concerns, we would appreciate it if you would reconsider your rating in light of the revised manuscript.

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

This paper addresses PDE-constrained optimization problems. The proposed framework combines a neural operator that approximates the underlying PDE dynamics with a proxy controller that predicts control actions. Experiments are conducted on voltage regulation, the heat equation, and Burgers' equation. The proposed method achieves competitive solution quality while significantly reducing inference time compared with several classical and learning-based control methods.

Critical strengths
  1. The paper is well-written and easy to follow. The experimental results are promising, with comparisons against both classical optimization methods and learning-based solvers.

  2. The paper addresses an important problem in PDE-constrained optimization. The significant reduction in inference time is particularly meaningful for real-time control applications.

Critical weaknesses
  1. The proposed method appears to be primarily a combination of DeepONet and a proxy optimizer trained through self-supervised primal-dual learning. While this integration is reasonable, the methodological novelty beyond these existing components is not sufficiently clear. The authors could provide a more detailed discussion to better highlight the novel contributions of the proposed framework.

  2. As evidenced by the experimental results (e.g., Table 19), the learned PDE surrogate does not fully capture the true system dynamics, despite achieving low prediction errors. These approximation errors can lead to substantially larger discrepancies in the final control performance when evaluated using the numerical PDE solver. This raises concerns about the reliability of optimizing control actions through the learned surrogate. I would be happy to raise the score if the above concerns are properly addressed.

Questions for authors
  1. The performance of the proposed framework appears to be highly dependent on the accuracy and capacity of the learned PDE surrogate, while the experiments mainly focus on relatively low-dimensional PDEs. How well does the framework generalize to more complex or higher-dimensional PDE systems? Is it compatible with alternative neural operator architectures?

  2. Why is only the terminal-state MSE reported as the primary metric for solution quality? According to the formulation in Eq. (1), the optimization objective also includes running costs and control regularization terms. Reporting the complete objective values would provide a more comprehensive evaluation of the optimization performance.

7 comments
AuthorsOpenReview ↗

Authors' response (1/4)

We thank the reviewer for the thoughtful assessment and for recognizing the practical relevance of fast inference for PDE-constrained control. We appreciate the concerns regarding methodological novelty, learned-dynamics reliability, architectural flexibility, and complete objective evaluation. We address each point below and highlight proposed manuscript additions in blue.

Weaknesses.

The proposed method appears to be primarily a combination of DeepONet and a proxy optimizer trained through self-supervised primal-dual learning. While this integration is reasonable, the methodological novelty beyond these existing components is not sufficiently clear. The authors could provide a more detailed discussion to better highlight the novel contributions of the proposed framework.

Response. We thank the reviewer for this comment. We acknowledge that neural operators and self-supervised primal–dual learning are established techniques, as already discussed in the Related Work section of the original manuscript. However, our contribution lies in their integration into a unified framework for PDE-constrained optimal control, enabling efficient approximation of control decisions while accounting for the underlying system dynamics and constraints.

Specifically, PDE-OP (i) learns a control-conditioned, time-discrete neural operator that directly predicts state transitions, (ii) trains a target-conditioned surrogate controller through differentiable rollouts of the learned dynamics without requiring optimal-control labels, and (iii) incorporates PDE dynamics, initial/boundary conditions, and operational constraints through a primal–dual training formulation.

Unlike SBTO and DeepONet MPC, PDE-OP avoids instance-specific numerical optimization at inference time. Compared with the closely related TI-DeepONet–DPC approach, PDE-OP directly learns state transitions rather than temporal derivatives advanced through numerical integration, and employs adaptive dual updates for constraint handling.

Thus, the novelty lies not in the individual components, but in their integration into a constraint-aware framework that enables fast, amortized solutions to PDE-constrained optimal control problems. We have further clarified these distinctions in the revised manuscript.

As evidenced by the experimental results (e.g., Table 19), the learned PDE surrogate does not fully capture the true system dynamics, despite achieving low prediction errors. These approximation errors can lead to substantially larger discrepancies in the final control performance when evaluated using the numerical PDE solver. This raises concerns about the reliability of optimizing control actions through the learned surrogate. I would be happy to raise the score if the above concerns are properly addressed.

Response. We agree that approximation errors in a learned dynamics surrogate can affect the quality of the resulting control decisions. Any learned surrogate is subject to approximation error, and for the nonlinear PDE dynamics considered here, the magnitude of that error can vary substantially with the control input. Consequently, low average trajectory prediction error on the original test distribution can mask larger errors under the control inputs selected by the controller. We investigate this issue on the Voltage benchmark by fine-tuning the dynamics surrogate on control-relevant trajectories and evaluating the resulting controls with an independent numerical PDE solver.

We generated additional numerical trajectories under control-relevant inputs and used them to fine-tune the pretrained Voltage operator. We then trained two controllers under matched experimental conditions, varying only the dynamics operator used during training, and evaluated their control actions using an independent numerical PDE solver.

On 128 unseen Voltage targets, the controller trained with the fine-tuned operator achieved a 63.97% lower mean physical objective than the controller trained with the original operator. This result shows that targeted fine-tuning of the dynamics surrogate can substantially improve downstream control performance on the Voltage benchmark. The experimental setting and results are provided below and in Appendix 7.11.1 of the revised manuscript.

Proposed manuscript addition

Appendix: Control-Relevant Operator Fine-Tuning and Physical Validation

X.1. Numerical data generation.

We evaluated the original Voltage operator under controller-generated controls and observed discrepancies between its predictions and the corresponding trajectories computed using the numerical Crank–Nicolson solver. This suggests that the operator may be less accurate for control actions that differ from those encountered during its initial training.

To improve its accuracy in these regions, we generated additional numerical trajectories using control inputs representative of those produced by the controller.

AuthorsOpenReview ↗

Authors' response (2/4)

In what follows, a raw control denotes the controller output before enforcing actuator bounds, while a projected control is obtained through the componentwise box projection

ujproj=min⁡(ujmax⁡,max⁡(ujmin⁡,ujraw))u_{j}^{\mathrm{proj}}=\min(u_{j}^{\max},\max(u_{j}^{\min},u_{j}^{\mathrm{raw}})).

For Voltage, this projection enforces the prescribed control bounds. We evaluate the state trajectories associated with the projected controls using the numerical PDE solver.

We considered five types of control inputs: raw controller outputs, box-projected controls, perturbed projected controls, smooth random controls, and oscillatory projected controls. Each control was applied to the independent numerical Crank–Nicolson (CN) solver to generate the corresponding state trajectory.

The resulting dataset consists of 320 training, 80 validation, and 80 held-out trajectories. The validation set was used for model selection, while the held-out set was reserved for evaluating the operator's prediction accuracy.

X.2. Operator fine-tuning.

Starting from the pretrained Voltage operator, we fine-tuned its parameters using the newly generated numerical trajectories together with previously available state-supervision data. This additional training aims to improve the operator's predictions for control inputs relevant to the optimization task, without changing the underlying model architecture.

Fine-tuning was performed using AdamW for 2,000 updates with a learning rate of 10−410^{-4}, across two data seeds.

X.3. Matched physical-control evaluation.

To assess the effect of operator fine-tuning on control performance, we trained two controllers: one using the original pretrained operator and another using the fine-tuned operator.

Both controllers were trained under the same conditions, including identical initialization, target streams within each seed, optimizer, 3,000-update budget, control space, and the same Voltage objective. In both cases, the corresponding dynamics operator was kept fixed during controller training.

We then evaluated the two controllers on the same 128 previously unseen Voltage targets. The predicted controls were projected using the componentwise box projection above and applied to the independent numerical CN solver.

The evaluation reports the numerical tracking MSE and the physical objective JJ, which includes terminal tracking error and control effort.

Training dynamicsNumerical MSE ↓\downarrowPhysical JJ ↓\downarrow
Original pretrained operator0.00230.00230.00230.0023
Fine-tuned operator0.0008\mathbf{0.0008}0.0008\mathbf{0.0008}

Table X: Matched Voltage controller training on 128 common unseen targets. Controls are projected and evaluated using an independent numerical PDE solver.

The controller trained with the fine-tuned operator achieved a 63.97% reduction in mean physical objective relative to the controller trained with the original operator. The reduction was computed from the unrounded measurements (paired difference −1.4483×10−3-1.4483\times10^{-3}; descriptive 95% bootstrap interval [−1.8334×10−3,−1.1179×10−3][-1.8334\times10^{-3},-1.1179\times10^{-3}]).

These results provide empirical evidence that targeted fine-tuning on control-relevant inputs can mitigate the impact of surrogate approximation error on downstream control performance. The exact feasibility guarantee applies to the control bounds enforced by the box projection.

Questions.

The performance of the proposed framework appears to be highly dependent on the accuracy and capacity of the learned PDE surrogate, while the experiments mainly focus on relatively low-dimensional PDEs. How well does the framework generalize to more complex or higher-dimensional PDE systems? Is it compatible with alternative neural operator architectures?

Response. To investigate the dependence of control quality on the learned dynamics, we conducted additional experiments varying the capacity and training-data budget of the Voltage operator.

We also evaluated the compatibility of PDE-OP with a Fourier Neural Operator (FNO) on a two-dimensional Heat-control problem. The FNO-based controller achieved a mean physical objective of 0.05510.0551 and tracking MSE of 0.04200.0420 on 40 unseen targets, with no observed control-bound violations.

For context, the original PDE-OP framework achieves a physical objective of 0.03700.0370 on the 2D Heat benchmark, compared with 0.03540.0354 for LMPC and Adjoint, 0.03900.0390 for DeepONet MPC, and 0.05930.0593 for SBTO. The FNO-based implementation achieves a physical objective of the same order of magnitude, demonstrating that the proposed controller-training framework can successfully accommodate alternative neural operator architectures.

AuthorsOpenReview ↗

Authors' response (3/4)

Together with the operator-capacity ablation, these results provide further evidence of the flexibility of PDE-OP, while broader scalability to more complex PDE systems remains an important direction for future work.

Proposed manuscript addition

Appendix: Operator Capacity and FNO Compatibility

Y.1. Operator capacity and data sensitivity.

We evaluate the influence of operator width, depth, and training-data size on Voltage control performance. Each configuration receives 10,000 operator updates and 5,000 controller updates, with results averaged over two repetitions on 128 unseen targets. All objectives are evaluated using numerical PDE execution.

Operator/data settingPhysical JJ ↓\downarrow
Base (width 128, depth 2)0.00150.0015
Width 640.00160.0016
Width 2560.00130.0013
Depth 10.00230.0023
Depth 30.0011\mathbf{0.0011}
Half training data0.00130.0013

Table Y1: Voltage operator capacity and training-data ablation.

The results indicate that increasing operator capacity can improve downstream control quality, with the depth-3 model achieving the lowest objective among the tested variants.

Y.2. FNO compatibility on 2D Heat.

We consider the two-dimensional reaction–diffusion system

∂ty=DΔy−β(y−yref)+αu\partial_{t} y = D\Delta y-\beta(y-y_{\mathrm{ref}})+\alpha u and u(t,x)=∑m=116cm(t)ϕm(x)u(t,x)=\sum_{m=1}^{16}c_m(t)\phi_m(x).

The system uses a 21 × 21 spatial grid and 40 Crank–Nicolson transitions. We adopt a 16-mode state representation and train an FNO to predict the next-state coefficients:

a^k+1=Fθ(ak,ck)\hat a_{k+1}=\mathcal F_\theta(a_k,c_k) and LFNO=E[∥Fθ(ak,ck)−ak+1CN∥22]\mathcal L_{\mathrm{FNO}} =\mathbb E\left[ \|\mathcal F_\theta(a_k,c_k)-a_{k+1}^{\mathrm{CN}}\|_2^2 \right].

The FNO is trained on 100 numerical trajectories, with 20 validation and 40 held-out trajectories per seed, using two training seeds. The operator receives 5,000 training updates, followed by 3,000 controller updates.

Rollout MSETerminal prediction MSE
2.8649×10−42.8649\times10^{-4}8.4415×10−48.4415\times10^{-4}

Table Y2: Held-out FNO prediction errors.

We evaluate the resulting feedback controller on 40 unseen Heat targets using the numerical PDE solver, with each control action computed from the current numerical state.

Physical JJ ↓\downarrowTracking MSE ↓\downarrowBox violations
0.05510.05510.04200.042000

Table Y3: Numerical FNO control evaluation on 40 unseen 2D Heat targets.

For context, the original PDE-OP implementation achieves a physical objective of 0.03700.0370 on the 2D Heat benchmark, compared with 0.03540.0354 for LMPC and Adjoint, 0.03900.0390 for DeepONet MPC, and 0.05930.0593 for SBTO. The FNO-based controller achieves a physical objective of the same order of magnitude, supporting the compatibility of PDE-OP with alternative neural operator architectures. While the experiment demonstrates this flexibility on a two-dimensional PDE, the 16-mode representation limits conclusions about scalability to more complex systems.

Why is only the terminal-state MSE reported as the primary metric for solution quality? According to the formulation in Eq. (1), the optimization objective also includes running costs and control regularization terms. Reporting the complete objective values would provide a more comprehensive evaluation of the optimization performance.

Response. We thank the reviewer for this question and agree that terminal MSE alone does not capture the complete optimization objective. The revised version of the manuscript now report the terminal, running, and control-effort contributions, together with the full objective evaluated using the numerical PDE solver, for PDE-OP and all baselines on the same 40 test targets per PDE.

Proposed manuscript addition

Appendix: Full-Objective Comparison

For each method, the numerical PDE solver evaluates

Jfull=Eterm+λErun+γEctrlJ_{\mathrm{full}} = E_{\mathrm{term}} + \lambda E_{\mathrm{run}} + \gamma E_{\mathrm{ctrl}},

with λ=0\lambda=0 for Voltage, λ=0.1\lambda=0.1 for Heat and Burgers, and γ=10−5\gamma=10^{-5} for all three problems. The following table reports the weighted objective components and inference runtime.

Reviewer_JPrMOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces PDE-OP, a framework for efficiently approximating solutions to PDE-constrained optimization problems. It employs a discrete-time neural operator and a surrogate controller, which predicts control actions and is trained through differentiable rollouts of the learned dynamics. It also adopts a learning framework to incorporate dynamics constraints, boundary conditions, and control constraints into the training objective to encourage satisfying the contraints.

Critical strengths
  • This paper introduces a framework that integrates learned PDE dynamics with control policy optimization, avoiding iterative optimization at inference time and enables fast control decisions.
  • The results show improvements in inference efficiency across several PDE-constrained optimization tasks while maintaining competitive solution quality, showing the practical potential of the proposed method.
Critical weaknesses
  • A soft-constraint approach does not guarantee that the predicted control actions satisfy all constraints.
  • It requires offline training. The overall computational advantage may be less significant than the reported speedups.
  • The novelty is not significant since combining learned neural operators with differentiable control policy optimization has been explored in prior work.
Questions for authors
  • Compared with existing approaches, what is the key novelty?
  • Could the authors report the constraint violations of all compared methods, as well as their optimality gaps?
  • How to determine whether the resulting control policies are feasible for practical applications?
  • Does the proposed method remain reliable under distribution shifts?
4 comments
AuthorsOpenReview ↗

Authors' response (1/3)

We appreciate the reviewer’s feedback and respond to each point below.

Weaknesses.

A soft-constraint approach does not guarantee that the predicted control actions satisfy all constraints.

Response. We thank the reviewer for this comment. We emphasize that the goal of PDE-OP is to provide fast, approximate solutions to PDE-constrained optimal control problems, as demonstrated by our numerical experiments. The proposed primal–dual training framework incorporates the problem constraints into the learning objective, allowing the controller to account for them when generating decisions. While this approach encourages constraint satisfaction, we acknowledge that it does not guarantee exact feasibility.

At the same time several class of constraint functions, including box constraints, which are common in PDE-constrained optimal control as they often represent physical limitations on control actions and actuators, can be enforced directly on the controller output through a simple hard projection. This guarantees satisfaction of the prescribed control-variable bounds without requiring additional numerical optimization. Our experiments confirm zero box-constraint violations after projection across all three PDE benchmarks.

PDEViolating cases before projectionViolating cases after projection
Voltage5/405/400/400/40
2D Heat39/4039/400/400/40
1D Burgers8/408/400/400/40

Table J1: Number of unseen test instances with at least one violation of the prescribed control-variable bounds, before and after projection (40 instances per PDE).

More generally, while exact feasibility can be efficiently enforced for certain classes of constraints, extending such guarantees to general nonlinear and potentially nonconvex constraints is considerably more challenging and would require solving more complex exact or approximate projections, potentially reducing the computational efficiency benefits that PDE-OP enables. Nevertheless, the proposed framework is compatible with more sophisticated projection-based correction methods that could improve or enforce constraint satisfaction. Investigating such extensions, while balancing feasibility guarantees and computational efficiency, represents an important direction for future work.

It requires offline training. The overall computational advantage may be less significant than the reported speedups.

Response. We would like to emphasize that offline training is a common requirement of learning-based approaches for PDE-constrained optimal control and is not unique to PDE-OP. In particular, learning-based baselines such as SBTO and DeepONet MPC also rely on pretrained neural surrogate models. However, unlike these methods, which still require iterative optimization at inference time, PDE-OP directly predicts control actions through a trained controller without solving an optimization problem at each iterative step. This is precisely the main computational advantage we aim to demonstrate: once trained, PDE-OP provides fast approximate control solutions without requiring instance-specific iterative optimization. The offline training cost can therefore be amortized over repeated inference queries. For transparency, the manuscript already reports the training times of both the dynamics operator and the surrogate controller for each PDE benchmark in Table 4.

The novelty is not significant since combining learned neural operators with differentiable control policy optimization has been explored in prior work.

Response. We acknowledge that neural operators and differentiable control policy optimization have been explored in prior work, as already discussed in Section 2 (Related Work) and Appendix 7.1 of the original manuscript. However, our contribution lies in their integration into a framework specifically designed to efficiently approximate solutions to PDE-constrained optimal control problems. In particular, PDE-OP (i) learns a control-conditioned, time-discrete neural operator that directly predicts state transitions, (ii) trains a surrogate controller through differentiable rollouts of the learned dynamics without requiring optimal control labels, and (iii) incorporates PDE dynamics, initial/boundary conditions, and control constraints through a primal–dual training formulation.

AuthorsOpenReview ↗

Authors' response (2/3)

Unlike SBTO and DeepONet MPC, PDE-OP avoids iterative optimization at inference time. Compared with the closely related TI-DeepONet–DPC approach, PDE-OP directly learns state transitions instead of temporal derivatives advanced through numerical integration, and employs adaptive dual updates to account for constraint violations during controller training. These distinctions, together with the experimental results, demonstrate the benefits of the proposed framework for fast approximate PDE-constrained optimal control. We have clarified these contributions and their relationship to existing work in the revised manuscript.

Questions.

Compared with existing approaches, what is the key novelty?

Response. As detailed in our response to Weakness 3, the key contribution of PDE-OP is a unified framework for PDE-constrained optimal control that integrates direct state-transition learning with self-supervised, constraint-aware controller training. This integration allows the controller to account for the underlying PDE dynamics and problem constraints while learning to generate control actions without optimal-control labels. Once trained, PDE-OP can efficiently approximate solutions to new problem instances without requiring iterative optimization at inference time. The novelty therefore lies in the proposed formulation and integration of the surrogate models for dynamics and optimization.

Could the authors report the constraint violations of all compared methods, as well as their optimality gaps?

Response. We thank the reviewer for this question. To address this, Appendix 9.1 of the revised manuscript reports a table which includes the control-bound violations and the objective gaps of all methods on the same 40 test targets for each PDE. We measure each gap relative to the numerical reference solution of the same problem instance, with all objectives evaluated using the numerical PDE solver.

Below we report Appendix, Section 9.1 of the revised manuscript.

Proposed manuscript addition

Appendix: Control-Constraint Violations and Numerical-Reference Objective Gaps

For each target ii, we construct a numerical reference by optimizing the control with the numerical PDE dynamics, without using a learned controller or learned surrogate:

uiref≈arg⁡min⁡u∈U  Jfull ⁣(yinum(u),u)u_i^{\mathrm{ref}} \approx \underset{u\in\mathcal U}{\arg\min}\; J_{\mathrm{full}}\!\left(y_i^{\mathrm{num}}(u),u\right) and Jiref=Jfull ⁣(yinum(uiref),uiref)J_i^{\mathrm{ref}}= J_{\mathrm{full}}\!\left(y_i^{\mathrm{num}}(u_i^{\mathrm{ref}}), u_i^{\mathrm{ref}}\right).

Here U\mathcal U is the admissible control set and yinum(u)y_i^{\mathrm{num}}(u) is the state trajectory computed by the numerical PDE model. For Voltage, we use bounded L-BFGS-B with exact Crank–Nicolson (CN) dynamics and numerical-model gradients; for 2D Heat, we solve a full-horizon quadratic program (QP) using the validated physical modal model; for Burgers, we use full-horizon L-BFGS-B with numerical dynamics and discrete-adjoint gradients.

For target ii and method mm, we then define the objective gap relative to this same-instance reference as

ΔJi,m=Ji,mfull−Jiref\Delta J_{i,m}=J_{i,m}^{\mathrm{full}}-J_i^{\mathrm{ref}} and ΔJ‾m=140∑i=140ΔJi,m\underset{m}{\overline{\Delta J}}=\frac{1}{40}\underset{i=1}{\sum}^{40}\underset{i,m}{\Delta J}.

The mean numerical-reference objectives are 0.00180.0018 (Voltage), 0.03540.0354 (2D Heat), and 0.10360.1036 (Burgers). The number of violating cases counts instances with at least one violation of the specified control-variable bounds. PDE-OP is evaluated after the hard box projection described in the response to Weakness 1.

PDEMethodViolating casesJ‾full↓\overline{J}_{\rm full}\downarrowΔJ‾ref↓\overline{\Delta J}_{\rm ref}\downarrow
VoltagePDE-OP0/400/400.00200.00200.00020.0002
Direct (FD)0/400/400.0018\mathbf{0.0018}2.0200×10−6\mathbf{2.0200\times10^{-6}}
Adjoint0/400/400.00180.00182.3400×10−62.3400\times10^{-6}
SBTO0/400/400.00480.00480.00300.0030
2D HeatPDE-OP0/400/400.03700.03700.00160.0016
LMPC0/400/400.0354\mathbf{0.0354}≈0\boldsymbol{\approx 0}
Adjoint0/400/400.03540.03542.1000×10−62.1000\times10^{-6}
DeepONet MPC0/400/400.03890.03890.00350.0035
SBTO0/400/400.05920.05920.02380.0238
BurgersPDE-OP0/400/400.11700.11700.01340.0134
SBTO0/400/400.1089\mathbf{0.1089}0.0053\mathbf{0.0053}
Adjoint0/400/400.11650.11650.01290.0129
NMPC0/400/400.11670.11670.01310.0131
DeepONet MPC0/400/400.11700.11700.01340.0134

Table J2: Control-bound violations, mean full objectives, and gaps relative to the numerical reference on 40 common targets per PDE.

AuthorsOpenReview ↗

Authors' response (3/3)

Overall, PDE-OP achieves objective values comparable to the numerical and learning-based baselines, while requiring substantially less inference time. Importantly, unlike the compared methods, PDE-OP generates control actions directly without solving an instance-specific optimization problem at inference time, demonstrating the computational advantage of the proposed learning-based framework.

How to determine whether the resulting control policies are feasible for practical applications?

Response. We assess the feasibility of PDE-OP's predicted control actions through two complementary steps. First, we apply a hard projection to ensure satisfaction of the prescribed control-variable bounds. Second, we evaluate the resulting controls using an independent numerical PDE solver, rather than relying solely on the learned surrogate. This allows us to assess the physical consistency and performance of the predicted controls under the governing PDE dynamics and initial/boundary conditions.

Across all three benchmarks, the projected control actions exhibit zero box-constraint violations on the 40 unseen test instances per PDE. Furthermore, we report PDE residuals and boundary-condition errors to quantify deviations from the governing physical constraints. Together, these evaluations provide a practical numerical procedure for assessing the feasibility and performance of the generated control actions under the considered PDE models.

Does the proposed method remain reliable under distribution shifts?

Response. To address this comment, we run additional experiments to evaluate pretrained PDE-OP controllers on test conditions that differ from training, without retraining on the shifted distributions. Specifically, we considered changes in target frequency, amplitude, and sharpness for Voltage; initial conditions, target profiles, and diffusion coefficients for 2D Heat; and viscosity and initial-condition amplitude for Burgers. The predicted controls are then assessed using a numerical PDE solver.

The following results demonstrate that PDE-OP maintains consistent performance under several variations in physical parameters and initial conditions, highlighting its robustness to the considered distribution shifts. More challenging target profiles lead to higher objective values, reflecting the increased difficulty of the corresponding control tasks.

Appendix: Numerical Evaluation Under Distribution Shifts

To evaluate the robustness of PDE-OP under distribution shifts, we consider test configurations that differ from those used during training, including variations in target profiles, initial conditions, and physical parameters. The pretrained controllers are evaluated on these unseen configurations without any additional training. For each setting, we report the mean full objective JfullJ_{\mathrm{full}} over 40 test instances, computed by applying the predicted controls to an independent numerical PDE solver.

PDETest conditionJ‾full↓\mathbf{\overline{J}_{\mathrm{full}}\downarrow}
VoltageIn-distribution (ID)0.00040.0004
Low-frequency targets0.01100.0110
High-frequency targets0.01830.0183
High-amplitude targets0.09760.0976
Sharp-step targets0.00710.0071
2D HeatIn-distribution (ID)0.05200.0520
Positive initial condition0.04980.0498
Negative initial condition0.05530.0553
Sharp target0.01900.0190
Diffusion ×0.75\times 0.750.05220.0522
Diffusion ×1.25\times 1.250.05210.0521
BurgersNominal (ν=0.03\nu=0.03, a=1.0a=1.0)0.08540.0854
ν=0.02\nu=0.02, a=1.0a=1.00.08110.0811
ν=0.04\nu=0.04, a=1.0a=1.00.09270.0927
ν=0.03\nu=0.03, a=0.5a=0.50.08300.0830
ν=0.03\nu=0.03, a=1.5a=1.50.08920.0892

Table J3: Numerical full-objective means under distribution shift conditions (40 targets per condition).

The results show relatively stable objective values under the considered Heat and Burgers parameter shifts, while more challenging Voltage target profiles lead to higher costs. These differences may also reflect changes in problem difficulty.

AuthorsOpenReview ↗

Authors' comment

Thank you again for your review. We wanted to follow up on our clarification of PDE-OP’s contribution, its offline training costs, and the distinction between exact enforcement of control bounds and numerical assessment of physical performance. We also reported objective gaps relative to numerical reference solutions and evaluated pretrained controllers under the specified distribution shifts. Please let us know if any part of these clarifications or numerical comparisons needs further explanation. If these responses address your concerns, we would appreciate it if you would reconsider your rating in light of the revised manuscript.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Authors' comment

Dear Reviewers,

Thank you for your thoughtful feedback. We have revised the manuscript, now titled Learning Fast Approximate Solvers for PDE-Constrained Optimal Control, and addressed your comments in our individual responses. Changes are highlighted in blue in the revised manuscript.

We clarified the scope of PDE-OP and its training pipeline and expanded the discussion of its contributions relative to prior work in Section 2 (Related Work). Table 4 reports the offline training times for the dynamics operator and surrogate controller. Appendix 9.1 reports control-bound violations and objective gaps relative to numerical reference solutions, and Appendix 7.11.1 presents the Voltage surrogate fine-tuning experiment and its numerical validation. The revised appendix also includes the full-objective comparison and additional experiments on distribution shifts, operator capacity, and compatibility with a Fourier Neural Operator.

Please see the revised manuscript accompanying this rebuttal. Our individual responses point to the sections and tables relevant to each comment.

Thank you for taking the time to review our work.

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.