acceptodds
Under review as a conference paper at ICLR 2027

Factorized Initial Noise Optimization with Diffusion and Flow Priors for Multi-Objective Scientific Inverse Design

Abstract

Scientific inverse design seeks structures in a vast design space that achieve a target property while satisfying domain-specific constraints. Generative models provide expressive priors over such design spaces, but existing inference-time methods steer generation toward either the target property or the constraints. We propose factorized initial noise optimization (FINO), which casts constrained inverse design as optimization over the initial noise of a pretrained generative model, and treats the reward and constraints as separate objectives. At each iteration, FINO first verifies whether the current iterate lies within the feasible set: if falls outside, then takes a constraint-restoration step; otherwise take a reward step toward the target property. This factorization confines reward optimization to the feasible region and requires no retraining, no modification of the sampling trajectory, and no constraint weights as trade-off. On topological photonic inverse design and molecule generation, FINO achieves low target-property error without sacrificing constraint satisfaction, and even robustly remains effective for target property values beyond the training distribution that of greatest scientific interest.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Initial Noise Optimization, Multi-objective Scientific Inverse Design, Diffusion and Flow models

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_782kOpenReview ↗
Rating
3(+2)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces FINO, an inference-time optimization framework for constrained inverse design using pretrained diffusion and flow models. FINO optimizes the initial noise by alternating between reward optimization and constraint-restoration steps, aiming to satisfy design constraints without manually balancing them against the reward. The method is evaluated on photonic inverse design and molecular property optimization.

Critical strengths
  • The problem is relevant to scientific inverse design, where optimizing target properties while respecting physical or chemical constraints is essential.

  • The method is relatively simple, applicable to pretrained diffusion and flow models, and requires no additional training.

  • The photonics experiments demonstrate promising results, particularly for out-of-distribution target properties. The use of independent FDTD simulations for evaluation strengthens these results.

Critical weaknesses
  • Limited methodological novelty: Initial noise optimization for reward-guided generation is well established, and separating objective optimization from constraint satisfaction is a classical idea in constrained optimization. The main contribution appears to be the particular combination of feasibility restoration, perturbation-based candidate selection, and reward optimization. It is unclear whether this constitutes a sufficiently substantial methodological advancement.

  • Missing relevant literature: Two particularly relevant papers are not discussed: Constrained Flow Optimization via Sequential Fine-Tuning for Molecular Design (Gutjahr et al., ICML 2026) and Source-Guided Flow Matching (Wang et al., ICLR 2026). These works overlap substantially with FINO's motivation and methodological ingredients. A clearer comparison is necessary to establish the novelty.

  • Reliance on differentiable rewards and constraints: The proposed method requires differentiable reward and constraint functions, which considerably limits its applicability to realistic scientific inverse-design problems. Many relevant properties and constraints are only accessible through non-differentiable simulations, expensive numerical evaluations, or experiments. Consequently, FINO often requires differentiable surrogate predictors, introducing additional modeling effort and potentially substantial approximation errors. This is especially concerning for the claimed out-of-distribution design capabilities, where surrogate reliability is inherently uncertain. The molecular experiments already demonstrate this limitation. Despite achieving zero constraint violations according to the surrogate, 12.8–26.6% of generated molecules violate the actual RDKit synthetic-accessibility constraint.

  • Insufficient baseline comparisons and ablations: The baselines use a common weighted reward–constraint objective, with weights selected based on InitNO, raising concerns about whether they are appropriately tuned for constraint satisfaction. Comparisons against standard constrained optimization procedures in the initial-noise space would be particularly informative. Furthermore, the contributions of individual components, especially perturbation-based candidate selection and feasibility restoration, are not adequately isolated.

  • Computational efficiency: The procedure requires repeated generator evaluations and gradient calculations, including additional rollouts for perturbation-based candidate selection. A comparison under matched computational budgets, including inference-time costs, would be valuable.

Questions for authors
  • How does FINO compare to standard constrained optimization algorithms applied directly to the initial-noise space, particularly under matched computational budgets?

  • How much of the improvement arises from separating reward and constraint updates, versus perturbation-based candidate selection?

  • How reliable are the differentiable surrogate predictors for out-of-distribution designs? Can the authors provide independent evaluations of the optimized molecular properties?

8 comments

Q1 Comparison across standard constrained optimization algorithms

A: We thank the reviewer for suggesting a comparison with standard constrained optimization methods applied directly to the initial-noise space.

To address this concern, we implemented an ALM-based optimization method in the initial-noise space. Similar to our CG+ALM baseline, the method first performs a reward update, followed by constraint projection using ALM. We conducted additional experiments by varying the reward-update parameter η\eta and the number of projection steps while fixing the inner and outer iteration counts on target bandgap value of 198meV.

Table R1. InitNO+ALM result on different experiment settings. Each entry reports MAE / constraint violation rate (%).

# Proj.η=1\eta = 1η=3\eta = 3η=10\eta = 10
5091.71 / 3528.56 / 354.88 / 55
1087.37 / 2542.49 / 702.25 / 60
583.75 / 3036.15 / 602.13 / 80

The results demonstrate a trade-off between target-property accuracy and constraint satisfaction. Increasing η\eta substantially improves MAE but generally increases constraint violations. Conversely, increasing the number of projection steps can improve constraint satisfaction at the expense of target-property accuracy.

These results suggest that directly applying ALM-based constraint projection in the initial-noise space does not necessarily achieve both high target-property accuracy and reliable constraint satisfaction. This further motivates FINO's feasibility-aware optimization strategy.

W1: Limited methodological novelty

A: First, we agree that initial noise optimization for reward-guided generation is well established ([1-2]). However, as demonstrated in InitNO results, directly extending initial noise optimization by combining the reward and constraints into a weighted objective does not reliably achieve both target-property accuracy and constraint satisfaction.

Second, we acknowledge that separating objective optimization from constraint satisfaction is a classical idea in constrained optimization. However, directly applying conventional constraint optimization methods to diffusion-based generation presents additional challenges. Applying constraint enforcement within the reverse diffusion trajectory modifies the pretrained sampling process, while applying conventional constrained optimization directly in the initial-noise space does not necessarily resolve the trade-off between reward maximization and constraint satisfaction. As demonstrated in Table 1 of the main paper, directly applying ALM-based projection in the initial-noise space improves feasibility under certain configurations but still exhibits substantial constraint violations or large target-property errors.

Our contribution is to bridge these two approaches by introducing factorized optimization in the initial-noise space, enabling reward maximization and constraint restoration to be handled separately without modifying the pretrained generative model or its sampling trajectory. Specifically, FINO integrates feasibility restoration, perturbation-based candidate selection, and reward optimization into a unified inference-time framework. Our experiments demonstrate that this formulation achieves competitive target-property accuracy while consistently satisfying the constraints across different scientific inverse-design tasks.

[1] Eyring et al. (2024). "ReNO: Enhancing One-step Text-to-Image Models through Reward-based Noise Optimization." NeurIPS 2024. [2] Tang et al. (2025). "Inference-Time Alignment of Diffusion Models with Direct Noise Optimization." ICML 2025.

W2: Missing relevant literature

A: We agree that both papers share important motivations with FINO, and incorporate them into the revised related work section to clarify our methodological contributions.

First, Constrained Flow Optimization (CFO) (Gutjahr et al., 2026) addresses a closely related problem of maximizing rewards while satisfying constraints, further highlighting the importance of multi-objective problem, also the target problem in this work. However, CFO approaches this problem through sequential fine-tuning of pretrained generative models, requiring task-specific model adaptation. In contrast, FINO operates entirely at inference time by optimizing the initial noise without updating the pretrained model parameters. Furthermore, while CFO is experimentally evaluated on synthetic benchmarks and molecular design, FINO demonstrates its applicability to both molecular generation and topological photonic inverse design.

Second, Source-Guided Flow Matching (SGFM) (Wang et al., 2026) shares our motivation of performing guidance through the source distribution while preserving the pretrained flow dynamics. SGFM provides a general framework for sampling from a modified source distribution, including optimization-based sampling methods closely related to existing initial noise optimization approaches such as D-Flow(Ben-Hamu et al. 2024). However, our methodological contribution goes beyond the standard approaches of initial noise optimization as shown in all these previous works. Rather, FINO specifically addresses constrained reward optimization by factorizing reward maximization and constraint satisfaction into separate updates, incorporating feasibility restoration and perturbation-based candidate selection to maintain constraint satisfaction while improving target-property accuracy.

We have cited both works in the related works of revised manuscript.

W4: Insufficient baseline comparisons and ablations

A: We thank the reviewer for raising concerns regarding baseline tuning and comparison fairness. Our original intention in using a shared weight combination was to highlight the practical difficulty of balancing reward maximization and constraint satisfaction through weighted objectives. However, we agree that method-specific hyperparameter tuning is necessary for a more comprehensive comparison. We therefore conducted additional weight sweeps and reran the experiments using independently selected baseline configurations.

First, we investigated the sensitivity of the fabrication constraint weight. We performed hyperparameter searches at a target bandgap of 0 meV, generating 20 designs for each configuration. As NN weight searching for 3 objectives costs N3N^3 we fix λK\lambda_K and λC3\lambda_{C_3} and sweep through λfab\lambda_{\text{fab}} only. Initially, we observed that classifier guidance (CG) required a substantially larger fabrication constraint weight to reduce constraint violations, as shown below.

Table R2. Fabrication constraint weight sweep of classifier guidance on target bandgap value of 0meV. Each entry reports MAE / constraint violation rate (%).

λfab\lambda_{\mathrm{fab}}11010210^210310^310410^410510^510610^6
CG0.32/540.28/540.33/440.34/140.54/80.68/40.71/4

Here, λK=10\lambda_K=10 and λC3=5\lambda_{C_3}=5 are fixed. Increasing λfab\lambda_{\mathrm{fab}} from 1 to 10510^5 reduces the violation rate from 54% to 4%, illustrating the sensitivity of constraint satisfaction to weight selection.

To improve the numerical scaling of the fabrication constraint, we reformulated its evaluator using the sum of pixel values with a threshold of 16, instead of the mean pixel value with a threshold of 0.25. These two formulations define the same feasibility condition for an 8×88\times8 patch, but differ in gradient magnitude. Using the rescaled evaluator, we heuristically search for the λK\lambda_K and λC3\lambda_{C_3} combination first and then sweep through λfab\lambda_{\text{fab}}

Table R3. Method-specific fabrication constraint weight sweep for BoN, CG, and SMC on target bandgap value of 0meV. Each entry reports MAE / constraint violation rate (%).

λfab\lambda_{\mathrm{fab}}151050100500
BoN2.24/02.93/03.46/03.46/03.46/03.46/0
CG0.26/350.35/100.44/50.40/00.38/0—
SMC1.69/451.80/401.86/352.95/153.33/52.96/5

The results demonstrate that different baseline methods exhibit different trade-offs between target-property accuracy and constraint satisfaction as the fabrication constraint weight increases. To prioritize constraint satisfaction, we selected the smallest fabrication weight at which the violation rate reached its minimum observed value and showed no further improvement with larger weights in our search. Based on this criterion, we selected (λK,λC3,λfab)=(10,5,1)(\lambda_K,\lambda_{C_3},\lambda_{\mathrm{fab}})=(10,5,1) for BoN, (10,5,50)(10,5,50) for CG, and (5,5,100)(5,5,100) for SMC.

For ALM, we independently investigated the number of inner iterations and the frequency of constraint projections during diffusion sampling.

Table R4. ALM hyperparameter sweep over inner iterations and constraint projections on target bandgap value of 0meV. Each entry reports MAE / constraint violation rate (%)

Iterations / # of Projections100010010
10.35 / 00.34 / 00.34 / 10
20.21 / 00.23 / 00.41 / 0
30.28 / 00.31 / 00.39 / 0

We selected two inner iterations and ten equally spaced projection steps for the subsequent experiments.

For InitNO, we repeated the weight sweep using the updated fabrication constraint calculation.

Table R5. InitNO weight sweep across reward and constraint coefficients on target bandgap value of 0meV. Each entry reports MAE / constraint violation rate (%).

λC3\(λK,λfab)\lambda_{C_3} \backslash (\lambda_K,\lambda_{\mathrm{fab}})(1,1)(1,5)(1,10)(5,1)(5,5)(5,10)(10,1)(10,5)(10,10)
150.61/036.30/547.76/013.48/2524.23/1019.97/06.96/507.99/3016.16/25
530.38/1050.39/027.60/013.68/513.51/3027.74/54.18/500.41/4022.45/35
1035.87/541.67/042.15/09.31/1517.88/1512.29/2011.73/5012.01/155.20/25

We selected (λK,λC3,λfab)=(10,5,5)(\lambda_K,\lambda_{C_3},\lambda_{\mathrm{fab}})=(10,5,5) for InitNO.

Reviewer_Kwi2OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

FINO addresses scientific inverse design by optimizing the initial noise of a pretrained diffusion or flow model, alternating between constraint restoration and target-property optimization. The method is evaluated on photonic inverse design and molecular generation.

Critical strengths
  1. The feasibility-first strategy is intuitive. It preserves the distinct roles of constraints and target-property objectives without retraining the generative model.

  2. The photonic experiments use independent FDTD simulations for evaluation, providing meaningful evidence beyond surrogate-predicted performance.

Critical weaknesses
  1. The formulation primarily addresses single-objective constrained optimization. I wonder how the combination of multiple constraints into the scalar constraint used by FINO. Another concern is that the theoretical analysis only establishes for constraint restoration, rather than convergence of the complete algorithm.

  2. The paper does not isolate the contributions of each component. Consequently, the specific benefit of the proposed factorization remains unclear.

  3. Appendix B tunes weights only for InitNO at the 198 meV target, and then transfers the selected setting to all baselines. My concern is that the compairison is not fair for other baseline. Also, it is not evident that the weighted-sum formulations is not useful at all.

  4. Although Algorithm 1 specifies the procedure, numerical settings for key parameters such as rollout lengths and optimization iterations are not reported.

Questions for authors
  1. Could you provide component ablations about how alternating updates contribute without perturbation, warm-up, and best-feasible-sample selection?
  2. Does FINO still hold its advantage with baseline under equivalent feasibility-based output selection and matched computational budgets?
  3. Could you clarify how are the multi photonic constraints combined into the scalar constraint, and explain why this setup is characterized as multi-objective?
  4. How is the lower-is-better SA score transformed to match the stated constraint F_c-\tau_c≥0 ?
5 comments

W1 & Q3: Problem Formulation Clarification

A: We thank the reviewer for pointing out the need for clarification.

First, our multi-objective characterization refers to the two distinct goals of reward maximization and constraint satisfaction. While the problem is formally expressed as single-objective constrained optimization, FINO treats these two goals separately due to their asymmetric roles: the reward should be maximized, whereas the constraints must be satisfied.

Second, in our photonic inverse-design experiments, we consider two constraints: C3 rotational symmetry and minimum fabrication size. We combine these constraints with equal weights during constraint restoration, since violating either constraint renders the generated design infeasible. For feasibility filtering and best-feasible-sample selection in Algorithm 1, we explicitly require both constraints to be satisfied, rather than relying solely on their aggregated value.

We revised Algorithm 1 and the problem formulation to explicitly clarify how FINO handles multiple constraints.

W2 & Q1: Component Ablation

A: We conducted additional experiments to investigate the contributions of warm-up initialization, random perturbation, and feasibility-aware sample selection.

First, we evaluate four configurations while retaining the alternating constraint-restoration and reward-update steps. The vanilla configuration performs these updates without warm-up initialization or random perturbation.

Table R1. Component ablation study at target 198 meV. Results report MAE (mean ±\pm std) and joint constraint violation rate over 50 generated designs.

ComponentMAE ↓\downarrowViol. (%) ↓\downarrow
Vanilla138.05±84.63138.05 \pm 84.630
+ Warmup135.16±87.78135.16 \pm 87.780
+ Perturbation1.49±1.551.49 \pm 1.550
+ Warmup + Perturbation1.07±0.741.07 \pm 0.740

The results show that random perturbation substantially improves target-property accuracy. We attribute this improvement to the exploration introduced by random perturbations, which can help optimization escape poor local optima.

As perturbation provides substantial improvements in MAE, we further investigate the effect of the number of perturbations.

Table R2. Ablation study on the number of perturbations at target 198 meV. Results report MAE (mean ±\pm std) and joint constraint violation rate.

PerturbationsMAE ↓\downarrowViol. (%) ↓\downarrow
0135.16±87.78135.16 \pm 87.780
121.54±40.4021.54 \pm 40.400
33.52±9.543.52 \pm 9.540
51.07±0.741.07 \pm 0.740
101.04±0.761.04 \pm 0.760

The results demonstrate that increasing the number of perturbations consistently improves target-property accuracy. These findings highlight the importance of perturbation-based exploration for improving reward optimization within FINO's factorized framework.

W4: Implementation Detail Report

A: For reproduction we report the entire detail for our experiments and its average sampling time of each method using 50 generated designs at a target bandgap of 54 meV.

Table R3. Hyperparameter configurations and average sampling time for baseline methods and FINO.

MethodλK\lambda_KλC3\lambda_{C3}λfab\lambda_{\mathrm{fab}}Diffusion StepsStep SizeIterationsParticles / PerturbationsProjectionsTime (s)
CG105110005–––16
ALM10––10005 / 0.23–10024
BoN1051100––256–60
SMC1051501–32–16
InitNO1051100.3100––24
FINO11110 / 100.3 / 0.5505–25

Q4: Constraint Formulation Clarification

A: For lower-is-better properties such as synthetic accessibility (SA), we transform the raw property into a signed constraint margin:

C(x)=τth−g(x)C(x) = \tau_{\mathrm{th}} - g(x)

where g(x)g(x) denotes the predicted SA score and τth=3.5\tau_{\mathrm{th}} = 3.5. Thus, C(x)≥0C(x) \geq 0 is equivalent to g(x)≤3.5g(x) \leq 3.5. This is consistent with our general formulation C(x)=Fc(x)−τcC(x) = F_c(x) - \tau_c by setting Fc(x)=−g(x)F_c(x) = -g(x) and τc=−τth\tau_c = -\tau_{\mathrm{th}}. Since FINO performs gradient ascent on C(x)C(x), the update naturally reduces the SA score when restoring constraint feasibility. We clarified this in Appendix D.

Thank you again for your sincere review. In the revision, we added our response addressing (1) problem formulation clarification, (2) component ablation of FINO, (3) implementation detail report, and (4) constraint formulation clarification. We hope these additions clarify our main contribution and address the questions you raised. We would greatly appreciate any further thoughts and opinions.

Dear Reviewer Kwi2,

We would like to kindly follow up, as the discussion period will close in less than a few hours. We hope our responses have addressed your concerns, and we would be happy to clarify any remaining questions in the time left. If you find our responses satisfactory, we would greatly appreciate it if you could take them into account in your final evaluation.

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

The paper proposes a new algorithm, called FINO (Factorized Initial Noise Optimization), for scientific inverse design. The setup involves a pre-trained diffusion/flow model, a reward and a constraint function, and the goal is to obtain a sample which maximizes the reward while satisfying the constraint. The algorithm involves two stages: a warmup phase where an initial Gaussian noise is brought into the feasible set (potentially involving resampling), followed by reward ascent and constraint restoration whenever the iterate leaves the feasible set. To my understanding, the claimed novelty of the paper is that it addresses both the reward (target) and the constraints separately, while most of the prior work addressed either the reward or the constraint, but not both. Experiments were performed on photonic inverse design and constrained molecule generation.

Critical strengths

(1) The main idea behind the algorithm, i.e., treating the reward and constraints differently, is well motivated, and the algorithm is presented clearly.

(2) The experiments indicate strong performance of the method compared to the baselines, showing that it is capable of generating samples that satisfy the constraints even when the target value is in the tail or out-of-distribution with respect to the training set.

Critical weaknesses

(1) My main concern is with the lack of theoretical contributions. Proposition 1 is the only such result, and as the authors themselves note, the second assumption (norm of the gradient being lower bounded) can fail to hold uniformly. This especially undermines the bound for the hitting time T_c, as the proof uses this assumption for all the iterates z_k. Additionally, this result only focuses on constraint satisfaction; there is no such analysis on the reward quality. For instance, it is not clear how the reward is affected during each constraint restoration step, and when selecting the largest margin candidate. It would be useful if the authors can address these points.

(2) The problem formulation in section 3 consists of a single reward function and a single constraint function, and the proposed method addresses this formulation. However, the photonic inverse design experiments use two constraints, and I could not find details on how their method was adapted to multiple constraints; could the authors elaborate on this? On a related note, can the authors explain how they treated the fabrication loss, which is not only non-differentiable, but also integer valued?

(3) While the experimental results showcase superior accuracy of the method over the baselines, there is no such comparison for average compute. I think this is important to include, because as noted by the authors, the warm up phase can involve resampling multiple times if the iterate does not reach the feasible set, especially in a high dimensional setting.

Questions for authors

Any clarifications/ experiments addressing points raised in weaknesses would be helpful.

4 comments

W1: Lack of theoretical contributions and analysis on the reward quality

A: We thank the reviewer for the thoughtful comment. We agree that Proposition 1 is a conditional result and that the lower-bound assumption on ∥∇c(zk)∥\|\nabla c(z_k)\| may not hold uniformly along all iterates.

However, FINO's final feasibility guarantee primarily follows from its algorithmic design rather than Proposition 1.

Regarding about the non-uniformality, our empirical results demonstrate that FINO can efficiently reach the feasible region despite this theoretical limitation, requiring only 0.82±1.140.82 \pm 1.14 warm-up optimization steps on average across 50 generated samples. Although these results do not theoretically guarantee termination, they demonstrate that the warm-up procedure requires relatively few optimization steps in practice.

Once warm-up identifies a feasible initialization, FINO maintains feasibility-first optimization algorithm to optimize the reward only within the feasible set. This solution is updated only when a new candidate satisfies all constraints and improves the reward. Therefore, even when intermediate iterates become infeasible or their reward degrades, the stored solution remains feasible and its reward is monotonically nondecreasing. Consequently, FINO guarantees the feasibility of its final selected output, provided that warm-up successfully identifies a feasible initialization.

Regarding the interaction between reward updates and constraint restoration, we provide the additional analysis on the reward quality on the representative optimization trajectory in the Appendix G. The results reveal two key observations:

(i) Constraint-restoration steps can either improve or degrade the reward, indicating that improving constraint satisfaction does not necessarily lead to better target-property accuracy.

(ii) Reward updates can move the current design either closer to or farther from the feasible region, demonstrating that reward maximization does not necessarily preserve constraint satisfaction.

These observations motivate FINO's factorized optimization strategy. Rather than requiring each update to simultaneously improve both objectives, FINO alternates between reward maximization and constraint restoration while preserving the best-feasible solution found so far.

We have revised the manuscript to clarify (1) the scope and limitations of Proposition 1, (2) the algorithmic feasibility preservation of the saved solution, and (3) the interaction between reward maximization and constraint restoration.

W2: Problem Forumulation Clarification

A: For simplicity we present the problem as a single constraint in Section 3, while our algorithm can be extended to multiple constraints. For example, our photonic inverse-design experiments consider two constraints: C3 rotational symmetry and minimum fabrication size. We combine these two constraints with equal weights during constraint restoration, as violating either constraint renders a generated design infeasible. The fabrication constraint was chosen as a practical fabrication guideline to avoid extremely small features that may be difficult to reproduce reliably during lithography and etching which requires at least 16 pixels. As directly counting binary pixels are non differentiable, we calculate the mean pixel value within an 8×88\times8 patch centered at each of the eight motifs. The fabrication constraint is defined as

Cfab(x)=min⁡j=1,…,8(164∑p∈Bjxp)−0.25,C_{\mathrm{fab}}(x)=\min_{j=1,\ldots,8}\left(\frac{1}{64}\sum_{p\in B_j}x_p\right)-0.25,

where BjB_j denotes the jj-th patch. The threshold of 0.250.25 corresponds to 16 pixels out of an 8×88\times8 region. This formulation allows the minimum size of each motif to be at least 16 pixels or larger.

We clarified these implementation details in the Appendix D.

W3: Computation Time Report

A: To address this concern, we measured the average sampling time of each method using 50 generated designs at a target bandgap of 54 meV. Table R1 reports the computational costs along with the corresponding hyperparameter configurations as reported in the Table 1 of main paper.

Table R1. Hyperparameter configurations and average sampling time for baseline methods and FINO.

MethodλK\lambda_KλC3\lambda_{C3}λfab\lambda_{\mathrm{fab}}Diffusion StepsStep SizeIterationsParticles / PerturbationsProjectionsTime (s)
CG105110005–––16
ALM10––10005 / 0.23–10024
BoN1051100––256–60
SMC1051501–32–16
InitNO1051100.3100––24
FINO11110 / 100.3 / 0.5505–25

Our method is computationally efficient, as the warm-up stage incurs minimal overhead, requiring only 0.82±1.140.82 \pm 1.14 optimization steps on average across 50 generated designs. Thus, FINO achieves reliable constraint satisfaction and competitive target-property accuracy with a computational cost comparable to other optimization-based methods. We included these results in the revised manuscript.

Thank you again for your sincere review. In the revision, we added our response addressing (1) the optimization trajectory of FINO, (2) problem formulation clarification, (3) computation time report on the photonic inverse design task. We hope these additions clarify our main contribution and address the questions you raised. We would greatly appreciate any further thoughts and opinions.

Dear Reviewer artN,

We would like to kindly follow up, as the discussion period will close in less than a few hours. We hope our responses have addressed your concerns, and we would be happy to clarify any remaining questions in the time left. If you find our responses satisfactory, we would greatly appreciate it if you could take them into account in your final evaluation.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.