Pith. sign in

REVIEW 4 major objections 9 minor 45 references

When Every Simulation Counts: Value-Based Reinforcement Learning for Accelerated Photonics Inverse Design

T0 review · 4 major / 9 minor · reviewed 2026-07-30 · grok-4.5

Pith's one-line read When full-wave laser simulations are scarce, only Dueling DQN reliably improves seven-variable PCSEL designs across matched starts.

desk verdict Credible matched bake-off: Dueling is the only value-based variant that lifts all four seeds under 83 FDTD calls; the reliability claim is real but under-powered and S-bound. read the letter →

arxiv 2607.23469 v1 pith:IDHUQACJ submitted 2026-07-26 physics.optics cs.AIcs.LGphysics.app-ph

classification physics.opticscs.AIcs.LGphysics.app-ph
keywords photonic-crystalsurface-emittinglasersinversedesignreinforcementlearningDeepQ-networkDuelingDQNsimulation-budget-awareoptimizationnanophotonics
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

Designing photonic-crystal surface-emitting lasers means tuning many coupled geometry and material parameters, and each candidate costs an expensive full-wave simulation. This paper asks which deep Q-learning tricks still help when the whole run is limited to about eighty simulator calls. Under one shared objective, one simulator, four matched starting seeds, and the same local-edit rules, it compares ordinary DQN with six variants. Only the dueling architecture—separating how good the current design neighborhood is from how good each individual edit is—improves every seed and produces better quality factor, wavelength accuracy, and upward power. The work matters because it turns a common lab constraint (few costly simulations) into a controlled test of which learning mechanisms actually allocate those calls well, instead of celebrating a single lucky final design.

What carries the argument

Dueling value–advantage decomposition: the network splits Q(s,a) into a state-value stream V(s) that judges whether the current design neighborhood is promising and an advantage stream A(s,a) that ranks the fourteen local edits, so scarce coupled FDTD feedback can be shared across actions rather than forcing an immediate full ranking of every edit.

What would settle it

Re-run the same four matched seeds and 83-call budget with an altered objective (different component weights or unclipped physical targets) or fabricate and measure the saved dueling endpoints; if another variant then wins consistently, or measured devices reverse the simulated ranking, the reliability claim for Dueling DQN fails.

Watch

Extended reading notes

Core claim

Under a matched 83-call FDTD protocol for a seven-variable PCSEL, Dueling DQN is the only tested value-based variant that improves all four seeds and leads on endpoint score, sample efficiency, and physical responses. Relative to first-evaluated designs its selected structures raise mean quality factor by about 26×, cut wavelength error by 64%, and raise upward power by 47%; versus baseline DQN they reach roughly 2.6× higher mean Q. Double DQN copies the baseline trajectories; Rainbow-lite can score high but depends strongly on the seed.

Load-bearing premise

The fixed hand-weighted scalar score used to rank designs is assumed to track real engineering quality well enough that winning on that score, with four seeds and one local-edit template, means the method is the most reliable for actual PCSEL design.

Editorial extensions

If this is right

  • Under tight full-wave budgets, prefer dueling heads before stacking distributional, prioritized-replay, or noisy extras for local-edit PCSEL search.
  • Matched seeds, shared initialization blocks, and process metrics (AUC, policy-selected high scores, raw Q/λ/power) become the standard way to credit algorithmic gains in simulation-driven photonics.
  • Double-style target fixes alone need not change the search path when the action space and replay are this small.
  • Combined Rainbow-style agents can unlock high upside but require explicit seed-dependence checks before deployment.
  • The same protocol can audit value-based RL on other costly Maxwell inverse-design tasks with discrete local edits.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If neighborhood value is the scarce signal, simpler dueling agents may beat heavier multi-component agents whenever simulator calls number in the tens rather than thousands.
  • Hand-built scalar utilities that gate Q by wavelength will systematically prefer balanced extraction over peak-Q cavities once confinement is clipped—designers should treat that ranking as a choice, not physics.
  • Extending the test to fabrication tolerances or continuous actions would show whether the dueling advantage survives the jump from idealized FDTD edits to buildable devices.
  • Incumbent-window local search plus replay may be doing as much work as the network variant; ablating the window schedule is a natural next control.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 9 minor

Summary. The manuscript compares baseline DQN against six value-based variants (dueling, C51 distributional, PER, Noisy, Double, and a Rainbow-lite combination) for seven-variable PCSEL inverse design under a tightly controlled protocol: shared Lumerical FDTD simulator, a fixed hand-weighted physics-aware score S (Eq. 9), an 83-call budget (20 shared random-init calls + 63 learned calls in three-edit incumbent-reset windows), and four matched seeds acting as experimental blocks. The central empirical claim is that Dueling DQN is the only variant that improves all four seeds (mean best score 87.51→93.60; min 84.65→88.70), with concurrent gains in AUC, Top-5, policy-selected high-score calls, and raw physics (Q 2.19→5.63×10^6, |Δλ| 0.86→0.26 nm, P 0.078→0.089); Double DQN exactly reproduces DQN trajectories (a clean null), and Rainbow-lite shows high upside but seed-dependent bimodality. Strengths include genuinely matched protocols, re-simulation audits of saved devices (Table 3), seed-level multi-metric reporting in the spirit of Agarwal et al. [1], disclosure that the strongest seed-1 endpoint arose from epsilon-random edits, and public code. The principal weaknesses are inferential: four seeds is thin support for a 'most reliable configuration tested' claim, no equal-budget non-RL baseline (e.g., random search) is reported, and the mechanism attribution (value–advantage decomposition) is partly undercut by the authors' own stochastic-path disclosure.

Significance. If the reliability claim holds, the paper provides (i) a rare controlled, equal-budget comparison of value-based RL variants on a real photonics problem, (ii) a reproducible public codebase, and (iii) a reporting template (matched seeds, shared init blocks, raw-metric audits, re-simulation checks) that is arguably more valuable than the specific winner, since it makes algorithmic attribution auditable in simulation-budget-limited inverse design. The physical gains for the dueling-selected devices (Q 2.19→5.63×10^6, |Δλ| 0.86→0.26 nm, P 0.078→0.089) are real and independently re-simulated, so the result is not purely an artifact of the scalar score. The main limitation on significance is inferential breadth: four seeds and one objective/template constrain how far 'most reliable configuration tested' can generalize, and no equal-budget non-RL baseline anchors the practical value of the RL machinery itself.

major comments (4)
  1. [§3.3 / §4.1 / Table 2] The headline reliability claim rests on four paired gains: +3.50, +9.63, +1.86, +9.39. A paired t-statistic on these is ≈3.2 at df=3 — borderline uncorrected, and not significant after any multiple-comparison adjustment across the seven variants. Moreover, per Table 5 the seeds also drive NumPy/PyTorch initialization and exploration draws, so 'matched' blocks share more than starting geometry. Since each run costs only 83 FDTD calls, I ask the authors either to add seeds (e.g., 8–12 total) or to report interval estimates in line with their own cited best practice [1] (stratified bootstrap CIs, IQM/median rather than mean over 4 runs) and to temper 'only variant that improves all seeds' and 'most reliable configuration tested' accordingly.
  2. [§4.2] The authors honestly report that the strongest endpoint (seed 1, 97.283 at call 64) arose from two epsilon-random edits and 'cannot be attributed directly to the dueling head.' Since 'improves all four seeds' is the paper's central selling point, the claim should be decomposable into policy-driven vs. exploration-driven gains. The reported 'policy ≥83' counts (7.25 vs 3.00) gesture at this, but a per-seed policy-only endpoint (best score attained by a greedy/noisy-greedy selected call) for every algorithm is needed to show the dueling advantage survives removal of stochastic discoveries.
  3. [§1 / §3.3 / §4] No equal-budget non-RL comparator is reported. For an optics readership, 'accelerated photonics inverse design' invites the question of whether any RL variant beats 83 calls of random search or a simple local hill-climber under the same S, budget, and incumbent-window protocol. The 20 forced-random calls logged in all 28 runs (560 evaluations) plus the shared init protocol may support an approximate random-search baseline at negligible extra simulation cost. Without this anchor, the practical-value framing in the abstract and Conclusion is unanchored even though the internal variant ranking is unaffected.
  4. [§4.2 (mechanism paragraph) / §4.3 (Double null)] The mechanism claim — that separating V(s) from A(s,a) is what helps — is weakened by the adjacent finding that Double DQN's altered targets never flip a single argmax across all 28 trajectories: value-estimate changes at this budget rarely change behavior. This raises the possibility that Dueling's gains come from architecture-induced tie-breaking or initialization rather than improved value learning. A load-bearing but cheap diagnostic: quantify how often the dueling and baseline Q-rankings disagree on matched states, and whether Dueling's greedy choices track the sign of the finite-difference responses in Fig. 6(b) better than DQN's do.
minor comments (9)
  1. [Abstract] The '(25.8×)' improvement is relative to the first evaluated designs, which are shared across all algorithms by construction; as phrased, casual readers may attribute it to Dueling specifically. Please clarify that this describes the shared init-to-endpoint trajectory.
  2. [Table 3 / Fig. 9(e)] All four re-simulated devices report θ = 0.841° exactly, and Fig. 9(e) is essentially flat. Please state the far-field angular grid resolution; if c_θ is effectively constant across all visited designs, the 'multi-physics' weighting in Eq. (9) is in practice a three-term objective, which deserves one sentence.
  3. [Table 6 (C51 projection)] The C51 support [0,5000] spans the theoretical return bound but rewards lie in [−10, ~100]; 51 atoms give ~98-unit spacing, and the −10 terminal penalty is clipped to the support edge. This may explain the distributional variant's poor showing and should be discussed as a fairness caveat for that arm.
  4. [Eq. (11)] Mean-centering the advantage over all 16 outputs includes the two masked width actions. This is harmless for the masked argmax (a per-state constant), but please say so explicitly to pre-empt confusion.
  5. [§2.2] The weights (60, 18, 8, 4, 10) and clip/log bounds are asserted without provenance. A one-dimensional sensitivity check — e.g., re-ranking endpoints under ±50% perturbation of the two largest weights from logged data — would cost no new simulations and would address transfer of the algorithm ranking beyond this specific S.
  6. [Fig. 2(b) caption] 'div.*' is described as a 'stricter visualization-only scale' but the scale is never defined; please define it in the caption of Fig. 8 as well.
  7. [Table 2] Seed 0's DQN best occurs at call 20, i.e., during random initialization — the learned phase added nothing in that block. Worth one explicit remark in §4.1, as it bears on how much of the baseline's performance is initialization.
  8. [Table 7(a)] 'First ≥83' for Rainbow-lite (21.5) is conditional on only 2/4 successful seeds; please annotate conditional means in the table to avoid cross-row misreading.
  9. [Conclusion / refs] Conclusion: 'endpoint, process, action-log and physical-response indicate' is missing a noun (e.g., 'analyses'). Reference [43] (sequence-modeling PCSEL design) is cited but never compared; one sentence on why it is out of scope would help.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: matched empirical RL comparison with a frozen shared objective, not a self-defining derivation.

full rationale

The paper’s load-bearing claim is an empirical ranking of DQN variants under a fixed PCSEL template, Lumerical FDTD simulator, 83-call budget, four matched seeds, and a hand-designed but frozen scalar utility S (Eq. 9) used as the RL reward. Algorithm differences are attributed to post-initialization query allocation, with Double DQN reproducing baseline trajectories as a controlled null and saved geometries re-simulated for consistency. Endpoint, AUC, policy, and raw FDTD responses (Q, λ, P, A_eff, θ) are reported separately; physical gains are not algebraically forced by the definition of S, nor fitted from the same quantities then relabeled as predictions. Self-citations to prior photonic RL work are background, not uniqueness theorems or load-bearing premises that forbid alternatives. There is no self-definitional identity, fitted-input-as-prediction loop, or renamed known law. Evaluating agents partly on the same S they optimize is standard RL practice and does not reduce the comparative result by construction. Circularity score 0.

Assumptions & free parameters 4 free parameters · 4 assumptions · 2 invented entities

The result is experimental, not axiomatic. Load-bearing inputs are standard RL/FDTD machinery plus paper-specific protocol choices: the scalar multi-physics score weights, discrete local-edit MDP, 83-call/4-seed matched design, and the assumption that FDTD outputs under the fixed template rank real design quality. No new physical entities are postulated.

free parameters (4)
  • Objective weights and normalizers in S (60, 18, 8, 4, 10; Q log clip bounds; λ Gaussian width 5 nm; P/A/θ clip scales) = 60 c_Q + 18 c_λ + 8 c_P + 4 c_A + 10 c_θ (Eq. 9)
    Hand-chosen hierarchy that defines the reward and primary ranking; not fitted online but arbitrary relative to physics units and decisive once Q and λ saturate.
  • Simulation budget and split (83 calls; 20 random init + 63 learned; 21 three-edit windows) = 83 FDTD calls; learn from call 21
    Defines the scarce-data regime in which “reliability” is measured; different budgets could reorder methods.
  • Network/optimization hyperparameters (MLP 8–80–120–80, γ=0.98, lr=1e-4, batch 16, target every 50, ε schedule, PER α/β, = See Appendix Tables 5–6
    Shared training choices that can interact with variant mechanisms; not swept in the paper.
  • Local edit step sizes (2.5 nm thicknesses/lattice; 0.005 index/hole-factor) = 2.5 nm / 0.005 as in §2.3 and Table 4
    Sets the MDP topology and which basins are reachable in 63 post-init steps.
assumptions (4)
  • domain assumption Classical Maxwell FDTD evaluations from Ansys Lumerical are adequate proxies for the PCSEL responses used in ranking (Q, λ, P, A_eff, θ).
    Entire optimization loop treats each solver return as ground truth; no fabrication or measurement validation (§Limitations).
  • domain assumption Q-learning / DQN bootstrap targets with replay and a target network are valid learning updates for this finite-horizon local-edit MDP.
    Baseline estimator and all variants inherit standard DQN theory (§3.1); not re-proved.
  • ad hoc to paper Matched random seeds, shared init trajectories, and identical action masks make cross-algorithm score differences attributable to value-learning mechanisms rather than start geometry.
    Core causal identification strategy of the bake-off (§3.3 Matched runs).
  • ad hoc to paper Masking width actions and fixing t4/w isolates a representative seven-variable PCSEL search without changing the reliability conclusion’s intended scope.
    Template reduction in §2.1; generalization beyond this parameterization is explicitly limited.
invented entities (2)
  • Physics-aware scalar score S with λ-gated Q term
    purpose: Provide a single fixed reward/ranking utility combining Q, wavelength, power, mode area, and divergence for RL and endpoint selection.
    Not a new physical observable; an engineering utility invented for the study (§2.2). Independent evidence is limited to consistency with stated design preferences and re-simulated raw fields.
  • Rainbow-lite agent (Double + dueling + PER + C51 + NoisyLinear, one-step only)
    purpose: Test component interactions without full original Rainbow (no multi-step returns).
    Paper-specific subset labeled Rainbow-lite (§3.2); results are about this construct, not canonical Rainbow.

how reviews work

0 comments
Cite this review

Pith. "Pith review of When Every Simulation Counts: Value-Based Reinforcement Learning for Accelerated Photonics Inverse Design." pith.science (2026). https://pith.science/paper/IDHUQACJ

@misc{pith2026260723469,
  author       = {Pith},
  title        = {Pith review of: When Every Simulation Counts: Value-Based Reinforcement Learning for Accelerated Photonics Inverse Design},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IDHUQACJ}},
  note         = {Machine review of arXiv:2607.23469}
}
read the original abstract

Photonic-crystal surface-emitting lasers (PCSELs) can combine high-power operation with narrow-divergence surface emission, but optimizing coupled parameters requires costly full-wave simulations. Deep Q-network (DQN) optimization can reuse simulated transitions to guide edits, yet which value-learning mechanisms remain reliable under tight simulation budgets is unknown. We address this gap by comparing baseline DQN and six value-based variants for a seven-variable PCSEL design under a shared objective, simulator, 83-call budget, and four matched initializations. Beyond endpoints, we analyze sample efficiency, policy behavior, and physical response to separate learning gains from favorable starts or exploratory jumps. Dueling DQN is the only variant to improve all four seeds. Relative to the first evaluated designs, its selected structures increase the mean quality factor () from to (), reduce wavelength error by 64%, and increase upward power by 47%; compared with baseline DQN, they achieve a higher mean under the same budget. Other variants yield no consistent improvement; Double DQN reproduces baseline trajectories, while Rainbow-lite shows high upside but strong seed dependence. These results identify Dueling DQN as the most reliable configuration tested for simulation-budget-limited PCSEL inverse design and provide a reproducible framework for attributing algorithmic gains in scientific optimization. The source code is publicly available at https://github.com/Longying-Wen/PCSEL-RL.

Figures

Figures reproduced from arXiv: 2607.23469 by the authors.

Figure 1
Figure 1. PCSEL geometry and controls. (a) Inset top view [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Matched endpoint and physical comparison. (a) [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Best-so-far score by seed. The vertical line ends the [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Shared-scale far-field intensity of (a) the highest [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Incumbent (a) 𝑄, (b) wavelength, and (c) upward power across calls (mean ± SD). The vertical line ends initial￾ization; dashed Double DQN overlaps DQN. Other metrics are in Appendix [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: Policy-sampled structure–response relations. (a) Visited displacement ranges; light whiskers denote min–max, dark [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 7
Figure 7. Figure 7: Saved near- and far-field profiles of two representative Dueling DQN endpoints: (a,b) the highest-scoring design (seed 1) [PITH_FULL_IMAGE:figures/full_fig_p011_7.png]
Figure 8
Figure 8. Figure 8: Seed-level physical-component differences relative to DQN. Panels (a)–(f) show dueling DQN, distributional DQN, [PITH_FULL_IMAGE:figures/full_fig_p013_8.png]
Figure 9
Figure 9. Figure 9: Supplementary incumbent physical trajectories completing Figure 5: (d) effective mode area and (e) far-field divergence. [PITH_FULL_IMAGE:figures/full_fig_p013_9.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

45 extracted references · 10 canonical work pages

  1. [1]

    Courville, and Marc G

    Rishabh Agarwal, Max Schwarzer, Pablo Samuel Castro, Aaron C. Courville, and Marc G. Bellemare. 2021. Deep Reinforcement Learning at the Edge of the Statistical Precipice. InAdvances in Neural Information Processing Systems, Vol. 34. 29304–29320

  2. [2]

    Takashi Asano and Susumu Noda. 2019. Iterative Optimization of Photonic Crystal Nanocavity Designs by Using Deep Neural Networks.Nanophotonics8, 12 (2019), 2243–2256. doi:10.1515/nanoph-2019-0308

  3. [3]

    Bellemare, Will Dabney, and Remi Munos

    Marc G. Bellemare, Will Dabney, and Remi Munos. 2017. A Distributional Perspec- tive on Reinforcement Learning. InProceedings of the 34th International Conference on Machine Learning (Proceedings of Machine Learning Research, Vol. 70). 449–458

  4. [4]

    Christiansen and Ole Sigmund

    Rasmus E. Christiansen and Ole Sigmund. 2021. Inverse Design in Photonics by Topology Optimization: Tutorial.Journal of the Optical Society of America B38, 2 (2021), 496–509. doi:10.1364/JOSAB.406048

  5. [5]

    Meire Fortunato, Mohammad Gheshlaghi Azar, Bilal Piot, Jacob Menick, Mat- teo Hessel, Ian Osband, Alex Graves, Vlad Mnih, Remi Munos, Demis Has- sabis, Olivier Pietquin, Charles Blundell, and Shane Legg. 2018. Noisy Net- works for Exploration. InInternational Conference on Learning Representations. arXiv:1706.10295

  6. [6]

    Peter Henderson, Riashat Islam, Philip Bachman, Joelle Pineau, Doina Precup, and David Meger. 2018. Deep Reinforcement Learning That Matters. InProceedings of the Thirty-Second AAAI Conference on Artificial Intelligence, Vol. 32. 3207–3214. doi:10.1609/aaai.v32i1.11694

  7. [7]

    Matteo Hessel, Joseph Modayil, Hado van Hasselt, Tom Schaul, Georg Ostrovski, Will Dabney, Dan Horgan, Bilal Piot, Mohammad Azar, and David Silver. 2018. Rainbow: Combining Improvements in Deep Reinforcement Learning. InPro- ceedings of the Thirty-Second AAAI Conference on Artificial Intelligence, Vol. 32. 3215–3222. doi:10.1609/aaai.v32i1.11796

  8. [8]

    Kazuyoshi Hirose, Yong Liang, Yoshitaka Kurosaka, Akiyoshi Watanabe, Takahiro Sugiyama, and Susumu Noda. 2014. Watt-Class High-Power, High-Beam-Quality Photonic-Crystal Lasers.Nature Photonics8 (2014), 406–411. doi:10.1038/nphoton. 2014.75

Show all 45 references
  1. [9]

    Hughes, Momchil Minkov, Ian A

    Tyler W. Hughes, Momchil Minkov, Ian A. D. Williamson, and Shanhui Fan. 2018. Adjoint Method and Inverse Design for Nonlinear Nanophotonic Devices.ACS Photonics5, 12 (2018), 4781–4787. doi:10.1021/acsphotonics.8b01522

  2. [10]

    Takuya Inoue, Yuya Seki, Shu Tanaka, Nozomu Togawa, Kenji Ishizaki, and Susumu Noda. 2022. Towards Optimization of Photonic-Crystal Surface-Emitting Lasers via Quantum Annealing.Optics Express30, 24 (2022), 43503–43512. doi:10. 1364/OE.476839

  3. [11]

    Kenji Ishizaki, Menaka De Zoysa, and Susumu Noda. 2019. Progress in Photonic-Crystal Surface-Emitting Lasers.Photonics6, 3 (2019), 96. doi:10. 3390/photonics6030096

  4. [12]

    Jensen and Ole Sigmund

    Jakob S. Jensen and Ole Sigmund. 2011. Topology Optimization for Nano- Photonics.Laser & Photonics Reviews5, 2 (2011), 308–321. doi:10.1002/lpor. 201000014

  5. [13]

    Anqing Jiang, Yoshie Osamu, and Liangyao Chen. 2020. Multilayer Optical Thin Film Design with Deep Q Learning.Scientific Reports10 (2020), 12780. doi:10.1038/s41598-020-69754-w

  6. [14]

    Jiaqi Jiang, Mingkun Chen, and Jonathan A. Fan. 2021. Deep Neural Networks for the Evaluation and Design of Photonic Devices.Nature Reviews Materials6, 8 (2021), 679–700. doi:10.1038/s41578-020-00260-1

  7. [15]

    Jones, Matthias Schonlau, and William J

    Donald R. Jones, Matthias Schonlau, and William J. Welch. 1998. Efficient Global Optimization of Expensive Black-Box Functions.Journal of Global Optimization 13 (1998), 455–492. doi:10.1023/A:1008306431147

  8. [16]

    Renjie Li, Ceyao Zhang, Wentao Xie, Yuanhao Gong, Feilong Ding, Hui Dai, Zihan Chen, Feng Yin, and Zhaoyu Zhang. 2023. Deep Reinforcement Learning Empowers Automated Inverse Design and Optimization of Photonic Crystals for Nanoscale Laser Cavities.Nanophotonics12, 2 (2023), 31...

  9. [17]

    Yong Liang, Chao Peng, Kyosuke Sakai, Seita Iwahashi, and Susumu Noda. 2011. Three-Dimensional Coupled-Wave Model for Square-Lattice Photonic Crystal Lasers with Transverse Electric Polarization: A General Approach.Physical Review B84, 19 (2011), 195119. doi:10.1103/PhysRevB.84.195119

  10. [18]

    Rodrigues, Kyu-Tae Lee, and Wenshan Cai

    Zhaocheng Liu, Dayu Zhu, Sean P. Rodrigues, Kyu-Tae Lee, and Wenshan Cai

  11. [19]

    Taigao Ma, Mustafa Tobah, Haozhu Wang, and L. Jay Guo. 2022. Benchmarking Deep Learning-Based Models on Nanophotonic Inverse Design Problems.Opto- Electronic Science1, 1 (2022), 210012. doi:10.29026/oes.2022.210012

  12. [20]

    Kudyshev, Alexandra Boltasseva, Wenshan Cai, and Yongmin Liu

    Wei Ma, Zhaocheng Liu, Zhaxylyk A. Kudyshev, Alexandra Boltasseva, Wenshan Cai, and Yongmin Liu. 2021. Deep Learning for the Design of Photonic Structures. Nature Photonics15 (2021), 77–90. doi:10.1038/s41566-020-0685-y

  13. [21]

    Momchil Minkov, Ian A. D. Williamson, Lucio C. Andreani, Dario Gerace, Be- icheng Lou, Alex Y. Song, Tyler W. Hughes, and Shanhui Fan. 2020. Inverse Design of Photonic Crystals through Automatic Differentiation.ACS Photonics7, 7 (2020), 1729–1741. doi:10.1021/acsphotonics.0c00327

  14. [22]

    Rusu, Joel Veness, Marc G

    Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Andrei A. Rusu, Joel Veness, Marc G. Bellemare, Alex Graves, Martin Riedmiller, Andreas K. Fidjeland, Georg Ostrovski, Stig Petersen, Charles Beattie, Amir Sadik, Ioannis Antonoglou, Helen King, Dharshan Kumaran, Daan Wierstra, ...

  15. [23]

    Piggott, Weiliang Jin, Jelena Vuckovic, and Alejandro W

    Sean Molesky, Zin Lin, Alexander Y. Piggott, Weiliang Jin, Jelena Vuckovic, and Alejandro W. Rodriguez. 2018. Inverse Design in Nanophotonics.Nature Photonics 12 (2018), 659–670. doi:10.1038/s41566-018-0246-9

  16. [24]

    Susumu Noda, Kyoko Kitamura, Tsuyoshi Okino, Daiki Yasuda, and Yoshinori Tanaka. 2017. Photonic-Crystal Surface-Emitting Lasers: Review and Introduction of Modulated-Photonic Crystals.IEEE Journal of Selected Topics in Quantum Electronics23, 6 (2017), 1–7. Art. no. 4900107. do...

  17. [25]

    DeLacy, John D

    John Peurifoy, Yichen Shen, Li Jing, Yi Yang, Fidel Cano-Renteria, Brendan G. DeLacy, John D. Joannopoulos, Max Tegmark, and Marin Soljacic. 2018. Nanopho- tonic Particle Simulation and Inverse Design Using Artificial Neural Networks. Science Advances4, 6 (2018), eaar4206. doi...

  18. [26]

    Piggott, Jesse Lu, Konstantinos G

    Alexander Y. Piggott, Jesse Lu, Konstantinos G. Lagoudakis, Jan Petykiewicz, Thomas M. Babinec, and Jelena Vuckovic. 2015. Inverse Design and Demonstra- tion of a Compact and Broadband On-Chip Wavelength Demultiplexer.Nature Photonics9 (2015), 374–377. doi:10.1038/nphoton.2015.69

  19. [27]

    Piggott, Jan Petykiewicz, Logan Su, and Jelena Vuckovic

    Alexander Y. Piggott, Jan Petykiewicz, Logan Su, and Jelena Vuckovic. 2017. Fabrication-Constrained Nanophotonic Inverse Design.Scientific Reports7 (2017),

  20. [28]

    Tom Schaul, John Quan, Ioannis Antonoglou, and David Silver. 2016. Priori- tized Experience Replay. InInternational Conference on Learning Representations. arXiv:1511.05952

  21. [29]

    Philipp-Immanuel Schneider, Xavier Garcia Santiago, Victor Soltwisch, Martin Hammerschmidt, Sven Burger, and Carsten Rockstuhl. 2019. Benchmarking Five Global Optimization Approaches for Nano-Optical Shape Optimization and Parameter Reconstruction.ACS Photonics6, 11 (2019), 27...

  22. [30]

    Adams, and Nando de Fre- itas

    Bobak Shahriari, Kevin Swersky, Ziyu Wang, Ryan P. Adams, and Nando de Fre- itas. 2016. Taking the Human Out of the Loop: A Review of Bayesian Optimization. Proc. IEEE104, 1 (2016), 148–175. doi:10.1109/JPROC.2015.2494218

  23. [31]

    Jasper Snoek, Hugo Larochelle, and Ryan P. Adams. 2012. Practical Bayesian Optimization of Machine Learning Algorithms. InAdvances in Neural Information Processing Systems, Vol. 25

  24. [32]

    Song, Akhil Raj Kumar Kalapala, Weidong Zhou, and Shanhui Fan

    Alex Y. Song, Akhil Raj Kumar Kalapala, Weidong Zhou, and Shanhui Fan. 2018. First-Principles Simulation of Photonic Crystal Surface-Emitting Lasers Using Rigorous Coupled Wave Analysis.Applied Physics Letters113, 4 (2018), 041106. doi:10.1063/1.5045486

  25. [33]

    Sapra, Jan A

    Logan Su, Dries Vercruysse, Jinhie Skarda, Neil V. Sapra, Jan A. Petykiewicz, and Jelena Vuckovic. 2020. Nanophotonic Inverse Design with SPINS: Software Architecture and Practical Considerations.Applied Physics Reviews7, 1 (2020), 011407. doi:10.1063/1.5131263

  26. [34]

    Sutton and Andrew G

    Richard S. Sutton and Andrew G. Barto. 2018.Reinforcement Learning: An Intro- duction(2 ed.). MIT Press

  27. [35]

    Allen Taflove and Susan C. Hagness. 2005.Computational Electrodynamics: The Finite-Difference Time-Domain Method(3 ed.). Artech House

  28. [36]

    Hado van Hasselt, Arthur Guez, and David Silver. 2016. Deep Reinforcement Learning with Double Q-Learning. InProceedings of the Thirtieth AAAI Conference on Artificial Intelligence, Vol. 30. 2094–2100. doi:10.1609/aaai.v30i1.10295

  29. [37]

    Ziyu Wang, Tom Schaul, Matteo Hessel, Hado van Hasselt, Marc Lanctot, and Nando de Freitas. 2016. Dueling Network Architectures for Deep Reinforcement Learning. InProceedings of the 33rd International Conference on Machine Learning (Proceedings of Machine Learning Research, Vo...

  30. [38]

    Christopher J. C. H. Watkins and Peter Dayan. 1992. Q-Learning.Machine Learning8 (1992), 279–292. doi:10.1007/BF00992698

  31. [39]

    Wiecha, Arnaud Arbouet, Christian Girard, and Otto L

    Peter R. Wiecha, Arnaud Arbouet, Christian Girard, and Otto L. Muskens. 2021. Deep Learning in Nano-Photonics: Inverse Design and Beyond.Photonics Research 9, 5 (2021), B182–B200. doi:10.1364/PRJ.415960

  32. [40]

    Kane S. Yee. 1966. Numerical Solution of Initial Boundary Value Problems In- volving Maxwell’s Equations in Isotropic Media.IEEE Transactions on Antennas and Propagation14, 3 (1966), 302–307. doi:10.1109/TAP.1966.1138693

  33. [41]

    Fountaine, and Aaswath P

    Christopher Yeung, Benjamin Pham, Zihan Zhang, Katherine T. Fountaine, and Aaswath P. Raman. 2024. Hybrid Supervised and Reinforcement Learning for Design Optimization of Nanophotonic Structures.Optics Express32, 6 (2024), 9920–9930. doi:10.1364/OE.512159

  34. [42]

    Masahiro Yoshida, Shumpei Katsuno, Takuya Inoue, John Gelleta, Koki Izumi, Menaka De Zoysa, Kenji Ishizaki, and Susumu Noda. 2023. High-Brightness Scalable Continuous-Wave Single-Mode Photonic-Crystal Laser.Nature618, 7966 (2023), 727–732. doi:10.1038/s41586-023-06059-8

  35. [43]

    Ceyao Zhang, Renjie Li, Cheng Zhang, Zhaoyu Zhang, and Feng Yin. 2024. Inverse Design of Photonic Crystal Surface Emitting Lasers is a Sequence Modeling Problem.arXiv preprint arXiv:2403.05149(2024). arXiv:2403.05149 [physics.optics] Appendix For readability, the supplementary...

  36. [1786]

    doi:10.1038/s41598-017-01939-2

  37. [2018]

    doi:10.1021/acs.nanolett.8b03171

    Generative Model for the Inverse Design of Metasurfaces.Nano Letters18, 10 (2018), 6570–6576. doi:10.1021/acs.nanolett.8b03171

Pith tools

Reviewed July 30, 2026 · model on record in the stance chip above.