pith. sign in

arxiv: 2604.07111 · v1 · submitted 2026-04-08 · ⚛️ physics.optics

Photon density of states engineering with generative inverse design for scalable 3D photonic metamaterials

Zesen Zhou , Jeevan Rois , Matias Kagias This is my paper

Pith reviewed 2026-05-10 17:30 UTC · model grok-4.3

classification ⚛️ physics.optics
keywords photon density of statesinverse designconditional generative adversarial networkphotonic metamaterialsholographic lithography3D nanoarchitecturesmetasurface diffraction
0
0 comments X

The pith

A conditional generative adversarial network inverts the nonlinear geometry-to-response mapping to produce 3D photonic metamaterials with targeted photon density of states.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper develops an end-to-end inverse design method that combines a data-driven surrogate model with a conditional generative adversarial network to create three-dimensional photonic metamaterials having specific photon density of states. Photon density of states controls light-matter interactions and therefore matters for applications such as photocatalysis and solar energy harvesting. Direct design is difficult because geometry and spectral response relate in a nonlinear, non-unique way. The surrogate model predicts frequency-resolved density-of-states spectra from metasurface diffraction parameters and lithographic thresholds, while the generative network proposes new parameter sets that realize desired spectral features. The resulting structures exhibit high local density of states over a broad normalized frequency range and outperform the original training dataset.

Core claim

We present an end-to-end inverse design framework for tailoring the pDOS of 3D photonic metamaterials fabricated via metasurface-based holographic lithography. A data-driven forward surrogate model predicts frequency-resolved pDOS spectra from metasurface diffraction parameters and lithographic thresholds. Inverse design is performed using a conditional generative adversarial network that generates candidate metasurface diffraction parameters for target pDOS features. 3D structures featuring high local pDOS were obtained across a broad normalized frequency range and consistently outperformed those in the original dataset. Structural analysis revealed that these high pDOS architectures fall 1

What carries the argument

The conditional generative adversarial network (cGAN) conditioned on target pDOS features, which proposes metasurface diffraction parameters that the surrogate model maps to high-density-of-states spectra.

Load-bearing premise

The surrogate model trained on the existing dataset accurately predicts pDOS for the new parameter sets generated by the cGAN, and those parameter sets produce structures that can be fabricated and measured without large simulation-to-experiment discrepancies.

What would settle it

Fabricate the cGAN-generated structures via holographic lithography, measure their actual frequency-resolved photon density of states, and compare the measured spectra to the surrogate-model predictions; large systematic mismatches would falsify the pipeline.

Figures

Figures reproduced from arXiv: 2604.07111 by Jeevan Rois, Matias Kagias, Zesen Zhou.

Figure 1
Figure 1. Figure 1: FIG. 1 [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: FIG. 2 [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: FIG. 3 [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: FIG. 4 [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: FIG. 5 [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
read the original abstract

The photon density of states (pDOS) governs fundamental light matter interactions and is a critical parameter for designing next generation light driven technologies such as photocatalysis and solar energy harvesting. Achieving a target pDOS in 3D nanoarchitected structures remains challenging due to the nonlinear and non unique relationship between geometry and spectral response. Here, we present an end to end inverse design framework for tailoring the pDOS of 3D photonic metamaterials fabricated via the scalable nanofabrication approach of metasurface-based holographic lithography. A data driven forward surrogate model is constructed to predict frequency resolved pDOS spectra from metasurface diffraction parameters and lithographic thresholds. Inverse design is performed using a conditional generative adversarial network (cGAN) that generates candidate metasurface diffraction parameters for target pDOS features. 3D structures featuring high local pDOS were obtained across a broad normalized frequency range and consistently outperformed those in the original dataset. Structural analysis revealed that these high pDOS architectures fall into two predominant structural categories with similar rotational symmetry characteristics. Our work establishes the first inverse design strategy for 3D photonic metamaterials fabricated via holographic lithography.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The paper presents an end-to-end inverse-design pipeline for 3D photonic metamaterials fabricated by metasurface-based holographic lithography. A data-driven surrogate model maps metasurface diffraction parameters and lithographic thresholds to frequency-resolved photon density of states (pDOS) spectra; a conditional GAN then generates candidate parameter sets targeting high local pDOS. The authors report that the generated 3D structures achieve high pDOS across a broad normalized-frequency range, consistently outperform the training dataset, and fall into two predominant structural families with similar rotational symmetry. The work is positioned as the first inverse-design demonstration for this scalable fabrication route.

Significance. If the surrogate predictions remain accurate on the cGAN-generated designs and the resulting parameter sets map to faithfully fabricated structures whose measured pDOS matches the simulated spectra, the approach would constitute a meaningful step toward scalable, data-driven engineering of 3D photonic metamaterials. The combination of a fabrication-aware surrogate with a generative model addresses a genuine inverse-design challenge in a high-dimensional, nonlinear design space, and the identification of recurring structural motifs could guide future manual or automated design rules.

major comments (3)
  1. [Abstract] Abstract and results: the central claim that the cGAN-generated structures 'consistently outperformed those in the original dataset' is stated without quantitative support (mean pDOS values, standard deviations, number of evaluated samples, or statistical tests). Without these metrics it is impossible to judge the magnitude or robustness of the reported improvement.
  2. [Methods / Results] Surrogate-model validation: the forward model is trained on metasurface diffraction parameters and lithographic thresholds, yet no hold-out error, ablation study, or out-of-distribution accuracy assessment is reported for the parameter sets produced by the cGAN. Because the generator is free to explore regions away from the training manifold, surrogate error on these novel designs directly affects the validity of the claimed pDOS gains.
  3. [Discussion] Fabrication fidelity: the pipeline assumes that the lithographic-threshold parameters supplied to the surrogate accurately capture the nonlinear exposure response of holographic lithography, but no experimental spectra or simulation-to-fabrication discrepancy analysis is provided. This leaves the weakest link in the end-to-end claim untested.
minor comments (2)
  1. [Abstract] Abstract contains several hyphenation and spacing inconsistencies ('end to end', 'data driven', 'frequency resolved').
  2. [Results] The two 'predominant structural categories' are described qualitatively; a quantitative metric (e.g., symmetry-order histogram or principal-component projection) would strengthen the structural-analysis claim.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for their constructive and detailed feedback, which has helped us clarify and strengthen several aspects of the manuscript. We address each major comment point-by-point below. Revisions have been made to incorporate quantitative metrics and additional validation where feasible, while clarifying the computational scope of the current work.

read point-by-point responses

  1. Referee: [Abstract] Abstract and results: the central claim that the cGAN-generated structures 'consistently outperformed those in the original dataset' is stated without quantitative support (mean pDOS values, standard deviations, number of evaluated samples, or statistical tests). Without these metrics it is impossible to judge the magnitude or robustness of the reported improvement.

    Authors: We agree that the original claim lacked the quantitative details needed for rigorous evaluation. In the revised manuscript, we have added a new Results subsection and Figure 4 that reports the following metrics for 200 cGAN-generated structures versus 1000 training-set structures: mean peak local pDOS of 1.48 (std 0.11) for generated designs versus 1.09 (std 0.14) for the dataset; a two-sample t-test yields p < 0.001. Similar statistically significant gains are shown across the normalized frequency range 0.2–0.8. These additions directly support the outperformance statement. revision: yes

  2. Referee: [Methods / Results] Surrogate-model validation: the forward model is trained on metasurface diffraction parameters and lithographic thresholds, yet no hold-out error, ablation study, or out-of-distribution accuracy assessment is reported for the parameter sets produced by the cGAN. Because the generator is free to explore regions away from the training manifold, surrogate error on these novel designs directly affects the validity of the claimed pDOS gains.

    Authors: We acknowledge the critical need to validate the surrogate outside the training distribution. The revised Methods section now includes: (i) hold-out test MAE of 0.047 on normalized pDOS spectra (n=800); (ii) an ablation study comparing MLP, CNN, and transformer-based surrogate architectures with performance tables; and (iii) targeted OOD evaluation on 150 cGAN-generated parameter vectors, yielding average MAE 0.061—only modestly higher than in-distribution error. These results are reported in a new supplementary table and confirm that surrogate predictions remain reliable for the generated designs. revision: yes

  3. Referee: [Discussion] Fabrication fidelity: the pipeline assumes that the lithographic-threshold parameters supplied to the surrogate accurately capture the nonlinear exposure response of holographic lithography, but no experimental spectra or simulation-to-fabrication discrepancy analysis is provided. This leaves the weakest link in the end-to-end claim untested.

    Authors: The referee correctly notes that the present study is a computational demonstration of the inverse-design pipeline; no new experimental pDOS spectra are reported. We have expanded the Discussion to explicitly list the assumptions underlying the lithographic-threshold model, quantify expected sources of discrepancy (e.g., resist nonlinearity, dose variation), and state that experimental validation of fabricated structures is planned as follow-on work. This revision clarifies the current scope without overstating the end-to-end claim. revision: partial

Circularity Check

0 steps flagged

No significant circularity in the inverse-design pipeline

full rationale

The paper presents a standard data-driven workflow: a forward surrogate is trained on existing metasurface diffraction and lithographic data to predict pDOS spectra, after which a cGAN is trained to generate new diffraction-parameter sets conditioned on target pDOS features. The claim that the generated structures 'outperformed those in the original dataset' is evaluated by applying the already-trained surrogate to the cGAN outputs and comparing the resulting scores to the original training-set pDOS values. This comparison does not reduce by construction to any fitted quantity inside the same model, nor does it rely on self-definitional equations, self-citation load-bearing uniqueness theorems, or ansatz smuggling. No equations or sections in the provided text equate a reported performance metric to a parameter that was itself optimized on that same metric. The pipeline therefore remains self-contained against external benchmarks and receives the default non-circularity finding.

Axiom & Free-Parameter Ledger

2 free parameters · 2 axioms · 0 invented entities

The central claim rests on two trained neural networks whose parameters are fitted to simulation data, plus the domain assumption that the surrogate remains accurate outside the training distribution and that generated designs are fabricable. No new physical entities are introduced.

free parameters (2)
  • cGAN and surrogate network weights
    All neural-network parameters are fitted during training on simulated pDOS data.
  • lithographic threshold parameters
    Threshold values used to map diffraction parameters to geometry are chosen or fitted as part of the forward model.
axioms (2)
  • domain assumption The forward surrogate model provides a sufficiently accurate mapping from metasurface diffraction parameters to frequency-resolved pDOS.
    Invoked to justify using the surrogate inside the inverse-design loop.
  • domain assumption cGAN-generated parameter sets correspond to physically realizable and measurable 3D structures.
    Required for the claim that high-pDOS structures were obtained.

pith-pipeline@v0.9.0 · 5500 in / 1466 out tokens · 97522 ms · 2026-05-10T17:30:53.382282+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

47 extracted references · 47 canonical work pages

  1. [1]

    S2(c)–(d)

    Training and testing performances are shown in Fig. S2(c)–(d). Inverse design model: FDTD-PSO simulation Metasurface patterns that reproduce the target diffrac- tion orders are obtained using a co-simulation framework that combines particle swarm optimization (PSO) with full-wave FDTD simulations (Fig. S3) [46]. The opti- mization process operates in an i...

    work page

  2. [2]

    Inhibited spontaneous emission in solid- state physics and electronics.Physical Review Letters58, 2059–2062 (1987)

    Yablonovitch, E. Inhibited spontaneous emission in solid- state physics and electronics.Physical Review Letters58, 2059–2062 (1987)

    work page 2059

  3. [3]

    Lodahl, P.et al.Controlling the dynamics of spontaneous emission from quantum dots by photonic crystals.Nature 430, 654–657 (2004)

    work page 2004

  4. [4]

    Bin, J.et al.Spontaneous-emission control by local den- sity of states of photonic crystal cavity.Chinese Physics B20, 024208 (2011)

    work page 2011

  5. [5]

    & Stobbe, S

    Lodahl, P., Mahmoodian, S. & Stobbe, S. Interfacing single photons and single quantum dots with photonic nanostructures.Reviews of Modern Physics87, 347–400 (2015)

    work page 2015

  6. [6]

    Zhao, J.et al.Free-standing, colored, polymer film with composite opal photonic crystal structure for efficient passive daytime radiative cooling.Solar Energy Mate- rials and Solar Cells251, 112136 (2023)

    work page 2023

  7. [7]

    Wang, X.et al.Engineered covalent triazine framework inverse opal beads for enhanced photocatalytic carbon dioxide reduction.Journal of Colloid and Interface Sci- ence689, 137244 (2025)

    work page 2025

  8. [8]

    Sun, K.et al.Full polarization and high coherence con- trol of thermal emissions via saddle-band dispersion en- gineering.Nature Communications16, 8393 (2025)

    work page 2025

  9. [9]

    G.et al.Nanophotonic engineering of far- field thermal emitters.Nature Materials18, 920–930 9 (2019)

    Baranov, D. G.et al.Nanophotonic engineering of far- field thermal emitters.Nature Materials18, 920–930 9 (2019)

    work page 2019

  10. [10]

    N.et al.Nanostructured fibers as a versatile photonic platform: radiative cooling and waveguiding through transverse anderson localization.Light: Science & Applications7, 37 (2018)

    Shi, N. N.et al.Nanostructured fibers as a versatile photonic platform: radiative cooling and waveguiding through transverse anderson localization.Light: Science & Applications7, 37 (2018)

    work page 2018

  11. [11]

    McPhedran, R.et al.Density of states functions for pho- tonic crystals.Physical Review E69, 016609 (2004)

    work page 2004

  12. [12]

    P., McPhedran, R

    Fussell, D. P., McPhedran, R. C. & de Sterke, C. M. Three-dimensional green’s tensor, local density of states, and spontaneous emission in finite two-dimensional pho- tonic crystals composed of cylinders.Physical Review E 70, 066608 (2004)

    work page 2004

  13. [13]

    M., Koenderink, A

    Johnson, P. M., Koenderink, A. F. & Vos, W. L. Ultra- fast switching of photonic density of states in photonic crystals.Physical Review B66, 081102 (2002)

    work page 2002

  14. [14]

    & Roberts, P

    Atkin, D., Russell, P., Birks, T. & Roberts, P. Photonic band structure of guided bloch modes in high index films fully etched through with periodic microstructure.Jour- nal of Modern Optics43, 1035–1053 (1996)

    work page 1996

  15. [15]

    & Loudon, R

    Barnett, S. & Loudon, R. Sum rule for modified sponta- neous emission rates.Physical Review Letters77, 2444– 2446 (1996)

    work page 1996

  16. [16]

    & Leung, K

    Yablonovitch, E., Gmitter, T. & Leung, K. Photonic band-structure - the face-centered-cubic case employing nonspherical atoms.Physical Review Letters67, 2295– 2298 (1991)

    work page 1991

  17. [17]

    & Vlasov, Y

    Norris, D. & Vlasov, Y. Chemical approaches to three- dimensional semiconductor photonic crystals.Advanced Materials13, 371–376 (2001)

    work page 2001

  18. [18]

    N., Harrison, M

    Campbell, M., Sharp, D. N., Harrison, M. T., Denning, R. G. & Turberfield, A. J. Fabrication of photonic crys- tals for the visible spectrum by holographic lithography. Nature404, 53–56 (2000)

    work page 2000

  19. [19]

    & Chichkov, B

    Serbin, J., Ovsianikov, A. & Chichkov, B. Fabrication of woodpile structures by two-photon polymerization and investigation of their optical properties.Optics Express 12, 5221–5228 (2004)

    work page 2004

  20. [20]

    T.et al.Enhanced photon management of thin-film silicon solar cells using inverse opal photonic crystals with 3d photonic bandgaps.Advanced Optical Materials1, 692–698 (2013)

    Varghese, L. T.et al.Enhanced photon management of thin-film silicon solar cells using inverse opal photonic crystals with 3d photonic bandgaps.Advanced Optical Materials1, 692–698 (2013)

    work page 2013

  21. [21]

    W., Richey, L

    Galusha, J. W., Richey, L. R., Gardner, J. S., Cha, J. N. & Bartl, M. H. Discovery of a diamond-based photonic crystal structure in beetle scales.Physical Review E77, 050904 (2008)

    work page 2008

  22. [22]

    Van Willenswaard, L. J. C.et al.Non-utopian optical properties computed of a tomographically reconstructed real photonic nanostructure.Optics Express32, 32028– 32047 (2024)

    work page 2024

  23. [23]

    A.et al.Optical properties of gyroid structured materials: From photonic crystals to metamaterials.Ad- vanced Optical Materials3, 12–32 (2015)

    Dolan, J. A.et al.Optical properties of gyroid structured materials: From photonic crystals to metamaterials.Ad- vanced Optical Materials3, 12–32 (2015)

    work page 2015

  24. [24]

    Peng, S.et al.Three-dimensional single gyroid photonic crystals with a mid-infrared bandgap.ACS Photonics3, 1131–1137 (2016)

    work page 2016

  25. [25]

    Siedentop, L.et al.Stealthy and hyperuniform isotropic photonic band gap structure in 3d.PNAS Nexus3, 383 (2024)

    work page 2024

  26. [26]

    K., Antonaglia, J., Dice, B

    Cersonsky, R. K., Antonaglia, J., Dice, B. D. & Glotzer, S. C. The diversity of three-dimensional photonic crys- tals.Nature Communications12, 2543 (2021)

    work page 2021

  27. [27]

    K., Nayak, S

    Cersonsky, R. K., Nayak, S. K. & Lee, S. H. Data-driven design rules for three-dimensional photonic crystals.ACS Omega11, 18400–18407 (2026)

    work page 2026

  28. [28]

    & Matsuo, S

    Kondo, T., Juodkazis, S., Mizeikis, V., Misawa, H. & Matsuo, S. Holographic lithography of periodic two- and three-dimensional microstructures in photoresist su- 8.Optics Express14, 7943–7953 (2006)

    work page 2006

  29. [29]

    M., Arbabi, E., Kwon, H

    Kamali, S. M., Arbabi, E., Kwon, H. & Faraon, A. Metasurface-generated complex 3-dimensional optical fields for interference lithography.PNAS116, 21379– 21384 (2019)

    work page 2019

  30. [30]

    Lee, J.et al.Proximity-field nanopatterning for high- performance chemical and mechanical sensor applications based on 3d nanostructures.Applied Physics Reviews9, 011322 (2022)

    work page 2022

  31. [31]

    Kagias, M.et al.Metasurface-enabled holographic lithog- raphy for impact-absorbing nanoarchitected sheets.Ad- vanced Materials35, 2209153 (2023)

    work page 2023

  32. [32]

    Zhang, C.et al.Low-loss metasurface optics down to the deep ultraviolet region.Light: Science & Applications9, 55 (2020)

    work page 2020

  33. [33]

    W., Sun, R., Lee, D., Portela, C

    Chen, W. W., Sun, R., Lee, D., Portela, C. M. & Chen, W. Generative inverse design of metamaterials with func- tional responses by interpretable learning.Advanced In- telligent Systems7(2025)

    work page 2025

  34. [34]

    & Huang, X

    Meng, F., Jia, B. & Huang, X. Topology-optimized 3d photonic structures with maximal omnidirectional bandgaps.Advanced Theory and Simulations1, 1800122 (2018)

    work page 2018

  35. [35]

    Wang, H.et al.Design of compact meta-crystal slab for general optical convolution.ACS Photonics9, 1358–1365 (2022)

    work page 2022

  36. [36]

    Men, H., Lee, K. Y. K., Freund, R. M., Peraire, J. & Johnson, S. G. Robust topology optimization of three- dimensional photonic-crystal band-gap structures.Optics Express22, 22632–22648 (2014)

    work page 2014

  37. [37]

    Y., Zhou, W., Huntington, M

    Hua, Y., Suh, J. Y., Zhou, W., Huntington, M. D. & Odom, T. W. Talbot effect beyond the paraxial limit at optical frequencies.Optics Express20, 14284–14291 (2012)

    work page 2012

  38. [38]

    Yuan, L. L. & Herman, P. R. Laser scanning holographic lithography for flexible 3d fabrication of multi-scale inte- grated nano-structures and optical biosensors.Scientific Reports6, 22294 (2016)

    work page 2016

  39. [39]

    M., Chan, C

    Ho, K. M., Chan, C. T. & Soukoulis, C. M. Existence of a photonic gap in periodic dielectric structures.Physical Review Letters65, 3152–3155 (1990)

    work page 1990

  40. [40]

    & Cao, L

    Jiang, Q., Jin, G. & Cao, L. When metasurface meets hologram: principle and advances.Advances in Optics and Photonics11, 518–576 (2019)

    work page 2019

  41. [41]

    G.et al.The control of shrinkage and ther- mal instability in su-8 photoresists for holographic lithog- raphy.Advanced Functional Materials21, 1593–1601 (2011)

    Denning, R. G.et al.The control of shrinkage and ther- mal instability in su-8 photoresists for holographic lithog- raphy.Advanced Functional Materials21, 1593–1601 (2011)

    work page 2011

  42. [42]

    & Dowling, J

    Florescu, M., Lee, H., Stimpson, A. & Dowling, J. Thermal emission and absorption of radiation in finite inverted-opal photonic crystals.Physical Review A72, 033821 (2005)

    work page 2005

  43. [43]

    Deng, W., Xu, Z., Wang, J. & Lv, J. Long short- term memory neural network for directly inverse design of nanofin metasurface.Optics Letters47, 3239–3242 (2022)

    work page 2022

  44. [44]

    Kingma, D. P. & Ba, J. Adam: A method for stochastic optimization. InInternational Conference on Learning Representations (ICLR)(2015)

    work page 2015

  45. [45]

    InAd- vances in Neural Information Processing Systems (NIPS) 10 (2014)

    Goodfellow, I.et al.Generative adversarial nets. InAd- vances in Neural Information Processing Systems (NIPS) 10 (2014)

    work page 2014

  46. [46]

    L., Hannun, A

    Maas, A. L., Hannun, A. Y. & Ng, A. Y. Rectifier non- linearities improve neural network acoustic models. In Proceedings of the 30th International Conference on Ma- chine Learning (ICML) Workshop(2013)

    work page 2013

  47. [47]

    Zhou, Y., Zeng, G. & Yu, F. Photonic crystal notch- filter design using particle swarm optimization and finite- difference time-domain method.Journal of the Optical Society of America B26, 849–853 (2009). DA T A A V AILABILITY The data that support the findings of this study are avail- able from the corresponding author upon reasonable re- quest. ACKNOWLE...

    work page 2009