acceptodds
Under review as a conference paper at ICLR 2027

Learning Latent Design Spaces for Efficient Optimization of Pixelated Electromagnetic Structures

Abstract

Pixelated electromagnetic (EM) structures offer substantial design flexibility, but their high-dimensional, combinatorial design spaces and costly full-wave simulations make target-driven optimization challenging. We investigate whether unsupervised representation learning can provide compact latent design spaces for efficient optimization of these structures. Engineering knowledge is encoded as geometric rules that generate large collections of layouts without EM response labels. We compare variational autoencoders, adversarial autoencoders, and an encoder-equipped Wasserstein generative model (WGAN-encoder) for learning low-dimensional representations of these geometric priors. Bayesian optimization then searches the learned spaces, using full-wave simulations to evaluate decoded candidates. We evaluate this pipeline on reflection metasurfaces (reflectarrays) and on-chip transformers, with four- and six-dimensional latent spaces, respectively. All three model families support effective downstream optimization, although convergence and final performance vary with the model and its regularization. A complete 16-cell reflection-phase library is obtained with 63 full-wave unit-cell evaluations and validated in a fabricated reflectarray, while transformer optimization improves the objective across all 16 target specifications. These results support learning from inexpensive geometric data as a practical approach to simulation-efficient optimization of pixelated EM structures across different representation models and device types.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Latent-Space Optimization, Bayesian Optimization, Electromagnetic Inverse Design

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_uMUAOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes an Pixelated electromagnetic structure design framework based on unsupervised representation learning and Bayesian optimization. Firstly, the author uses the engineering geometry prior to generate a large number of unlabeled structural layouts that meet specific geometric constraints without electromagnetic simulation, and uses VAE, AAE and WGAN-encoder to learn the low dimensional potential design space. Then, when the trained generation model is fixed, the latent variable is searched by Bayesian optimization, which is decoded into a specific structure, and the optimization is guided by the full wave electromagnetic simulation results. Finally, the author validates the framework on two design issues: reflectarray and on-chip transformer.

Critical strengths
  1. This paper presents a set of electromagnetic structure optimization process with practical application value. By using the engineering geometry prior to generate a large number of unlabeled legal layouts, it means that expensive full wave electromagnetic simulation data are not required in the learning phase, and the dependence on the training data with physical response labels is reduced.
  2. The author evaluates three kinds of representation learning methods on two design tasks, and verifies the effectiveness of the framework.
  3. The author has made a real reflection array. The measured pattern is consistent with the expected beam direction, which shows the application potential of this method in the design of practical electromagnetic devices.
Critical weaknesses
  1. This paper is quite similar to the existing work. The design framework based on unsupervised representation learning and Bayesian optimization has been widely used[1][2][3].

    The author believes that their difference lies in the use of geometric processing rather than finding the closest layout in the library[1]. However, geometric processing itself is not a new method. And the author has not proved that direct processing is superior to library matching. Furthermore, geometric processing can also be regarded as a huge implicit Library (projecting the generated shape onto the manifold formed by the legal shape)

[1] Tian Z, Yang Y, Zhou S, et al. High-dimensional Bayesian optimization for metamaterial design. MGE Advances. 2024; 2(4):e79. https://doi.org/10.1002/mgea.79

[2] R. Marzban, H. Abiri, R. Pestourie, and A. Adibi, “ HiLAB: A Hybrid Inverse-Design Framework.” Small Methods 9, no. 11 (2025): e00975. https://doi.org/10.1002/smtd.202500975

[3] Fan, Lei, and Zhongqing Su. "An efficient inverse design framework of manufacturable phononic metaplates via combining generative deep learning and Bayesian latent space exploration." Mechanical Systems and Signal Processing 243 (2026): 113677.

  1. Lack of in-depth exploration of the framework.

    a. lack of simple key optimization baselines

         Such as latent space + Random Search, PCA(t-SNE) + BO.
         These simple baselines can verify whether the latent space is effective and to what extent.
    

    b. lack of sufficient analysis of the learned latent design space.

     The author did not discuss what kind of latent space is more suitable for downstream tasks, only tested three existing representation learning methods, and did not design new representation learning methods. 
    
     A feasible example: a representation learning method is designed theoretically rather than experimentally, which is very suitable for Bayesian Optimization rather than other optimization methods (different optimization methods may adapt to different representation learning methods)
    
Questions for authors

I think it is more suitable for IEEE journals on electromagnetic computing, microwave engineering or antenna design, rather than ICLR conferences with machine learning method innovation as the main evaluation standard.

IEEE Journal on Multiscale and Multiphysics Computational Techniques

IEEE Transactions on Microwave Theory and Techniques

or conference

IEEE AP-S/URSI 2027

EuCAP 2027

Reviewer_ei36OpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents a workflow for optimizing pixelated electromagnetic structures using geometric priors, unsupervised representation learning, and Bayesian optimization (BO). Procedural rules generate layouts without electromagnetic response labels; a learned decoder then parameterizes the candidates evaluated by a full-wave solver. The applications cover reflectarray unit cells and on-chip transformers, including a fabricated reflectarray.

Critical strengths
  1. The reflectarray study provides an end-to-end engineering demonstration: a 16-target phase library is obtained with 63 cumulative unit-cell evaluations, including initialization, and used in a fabricated array with a measured peak realized gain of 25.57 dBi at 10 GHz. The array-level validation is separate from this unit-cell budget.

  2. The comparison of VAE, AAE, and Wasserstein variants, together with the supplementary AAE phase library, provides useful evidence that the workflow can operate with multiple representation models.

Critical weaknesses
  1. Limited methodological novelty. Latent-space optimization and the combination of adversarial autoencoders with Bayesian optimization are established ideas. Section 2 acknowledges relevant precedents, including AAE-based BO in Zhu et al. (2022) and learning from unlabeled geometry in Tian et al. (2024). The present work's distinguishing elements appear to be the specific geometric priors, the use of newly decoded layouts for full-wave evaluation, the organization of the workflow, and the device-level validation. These are useful engineering contributions, but the paper does not clearly establish a substantial methodological advance beyond the existing components. The model comparisons show that representation choice matters in the reported setting; they do not by themselves identify a broadly applicable new principle for representation learning or optimization. A sharper account of what is technically new, and evidence isolating its benefit, would strengthen the contribution.
  2. Missing baselines for quantifying simulation efficiency. The reported 63 evaluations establish the cost of one successful cumulative phase-library run, but do not quantify an efficiency advantage over alternative design strategies. Figure 4 and Table 3 provide useful five-seed downstream comparisons among VAE, AAE, and Wasserstein regularization variants within the latent BO pipeline. These comparisons do not determine whether learning a latent space improves over direct search, whether BO improves over random sampling within that space, or how much benefit comes from the geometric rules and cross-target reuse of simulations. The 2^D count in Section 1 describes the size of the unconstrained binary space, not the evaluation cost of a practical optimizer. By the 2^D count itself cannot establish an efficiency advantage over conventional, non-exhaustive design workflows. Useful baselines would include other suitable conventional design workflows. Even though section 5 explicitly leaves comparisons with direct pixel search under matched budgets to future work, it further worsen the current issues on lack of contribution.
  3. Uncharacterized restriction of the reachable design space. Equation (1) searches the rule-constrained design family, whereas Equation (4) searches only layouts reachable through the fixed decoder and geometric postprocessing. This restriction may make optimization easier, but it can also exclude high-performing structures that satisfy the original rules. The four- and six-dimensional latent spaces are used without a latent-dimension ablation or a systematic assessment of how the breadth of the geometric prior affects target coverage. Even though a formal global-optimality guarantee is not always required, empirical evidence is somewhat needed that the selected representations preserve enough of the relevant design space to justify the reduction.
  4. Unresolved causes of the unsuccessful designs. The transformer results contain substantial target mismatches despite an improvement in the objective for all 16 cases. The current experiments do not distinguish physical infeasibility under the specified device constraints from limited geometric-prior coverage, information lost through latent compression, or an insufficient search budget.
Questions for authors
  1. What methodological advance should distinguish this work from prior latent-space design methods, including AAE-based BO? Which experiment isolates the benefit of that advance beyond the use of established components?

  2. The 2^D design-space count does not imply that practical inverse-design algorithms require exhaustive enumeration. What are the appropriate conventional, non-exhaustive baselines for the two applications, and how does the proposed workflow compare under matched physical constraints and simulation budgets?

  3. How sensitive are target coverage, final objective values, and simulation requirements to latent dimension and geometric-prior breadth? Is there evidence that known high-performing layouts remain reachable after encoding, decoding, and postprocessing?

  4. For the difficult transformer targets, can the authors separate search-budget limitations from representation or prior limitations? Are physical bounds or independently obtained designs available to clarify which targets are feasible?

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

The paper investigates how to reduce the simulation cost of optimizing pixelated electromagnetic structures. It generates unlabeled layouts from engineering rules, learns compact representations using VAE, AAE, or WGAN-based models, and searches the resulting latent spaces with Bayesian optimization. The approach is evaluated on reflectarray unit cells and on-chip transformers. The main results include a 16-cell reflection-phase library obtained with 63 full-wave evaluations, a fabricated reflectarray, and reported objective improvements across 16 transformer specifications.

Critical strengths

Learning from geometry alone avoids collecting expensive simulation labels for representation training. The two device applications show that the workflow can accommodate different geometric constraints. The fabricated reflectarray provides meaningful physical validation beyond simulation. The five-seed comparison also offers useful evidence about the effect of representation choice.

Critical weaknesses
  1. The central efficiency benefit is not established against suitable baselines. Comparing several learned representations does not show whether learning improves on random sampling or direct optimization using the same geometric prior and simulation budget. The 63-evaluation result demonstrates feasibility, but the saving attributable to the learned space remains unclear.

  2. The methodological novelty is limited. The main contribution is an engineering pipeline combining established representation models and Bayesian optimization. The paper does not clearly distinguish its learning contribution from the latent design approaches already discussed in section 2.

  3. The transformer evidence is incomplete. The objective weights, achieved quality factors, and initial/final objective values are not reported, so the claimed improvement across all 16 cases cannot be checked. Several targets are also missed substantially—for example, a requested inductance of 120 pH produces 51.6 pH—leaving the demonstrated target coverage limited.

Questions for authors
  1. How does the method compare with sampling or direct optimization using the same geometric prior, simulation budget, and reuse of observations across targets?
  2. What are the transformer objective weights, achieved quality factors, initial/final objectives, and actual evaluation counts for each case?
  3. For the substantially missed transformer targets, is there evidence that suitable designs exist within the geometric prior and remain accessible through the learned decoder? This would help distinguish limitations of the prior, representation, and search.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.