Pith. sign in

REVIEW 2 major objections 1 minor 21 references

LLM agent designs silicon directional coupler to 0.0017 error in 50/50 split by correcting fixed excess length

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.3

2026-06-26 10:07 UTC pith:5JCZDKED

load-bearing objection The paper shows an LLM agent can steer existing solvers to a 50/50 coupler in a 2D model by measuring and folding in a constant excess length offset. the 2 major comments →

arxiv 2606.22493 v1 pith:5JCZDKED submitted 2026-06-21 physics.optics cs.AI

An LLM-Orchestrated Agent for Directional-Coupler Design with Self-Consistent Eigenmode and FDTD Validation

Saumya Biswas , Amrit De , Md Tauhidul Islam This is my paper
classification physics.optics cs.AI
keywords directional couplerLLM-orchestrated agenteigenmode solverFDTD validationsilicon-on-insulator50/50 splittercoupling coefficienteffective-index model
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper demonstrates an LLM that orchestrates the design of a 2x2 directional coupler on silicon-on-insulator by proposing gap values and using deterministic solvers for the physics. An eigenmode solver calculates the coupling coefficient kappa, and FDTD validates the response, both in a 2D effective-index model that keeps them consistent. The key finding is that any difference is a single constant phase offset from a fixed excess coupling length of 2.837 micrometers that stays the same even when kappa changes by a factor of two. By folding this offset into a length correction loop, the agent produces a design with FDTD cross fraction of 0.498 for a target of 0.5. The LLM can deliver suitable designs across multiple attempts within this self-consistent model.

Core claim

The LLM-orchestrated agent proposes gap values for a symmetric phase-matched directional coupler and, after identifying a fixed excess coupling length L_extra of 2.837(11) micrometers that causes a constant phase offset, applies a closed-loop length correction to achieve an FDTD-measured cross fraction of 0.498 against the 0.500 target with a residual of 0.0017, while remaining self-consistent in the 2D effective-index model.

What carries the argument

The LLM design agent that proposes candidate gaps and applies corrections based on the invariant excess coupling length identified from eigenmode and FDTD comparisons.

Load-bearing premise

The 2D effective-index reduction produces eigenmode and FDTD results that differ only by a single constant phase offset from a fixed excess coupling length invariant across kappa values.

What would settle it

Simulating designs with different kappa values after applying the length correction and observing whether the cross fraction remains within 0.002 of 0.5 would confirm or refute the invariance of the excess length.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • The agent achieves an FDTD cross fraction of 0.498 for the target 0.500 splitter.
  • The excess coupling length remains invariant across a factor-of-two range in the coupling coefficient kappa.
  • The design results are self-consistent within the 2D effective-index model.
  • The LLM succeeds in delivering suitable designs over multiple attempts.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • This approach may allow similar agents to design other phase-matched photonic devices using available analytical results.
  • The constant offset in the 2D model could enable reliable use of reduced-dimensional simulations for full 3D structures with simple adjustments.
  • Extending the agent to optimize multiple parameters simultaneously might handle more complex coupler designs.
  • Experimental fabrication and measurement would test if the reported accuracy holds beyond simulation.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 1 minor

Summary. The manuscript presents an LLM-orchestrated agent for designing a symmetric phase-matched 2×2 directional coupler on SOI. The agent proposes gap values while a frequency-domain eigenmode solver extracts the coupling coefficient κ and an independent FDTD solver validates the power splitting; both operate on the same 2D effective-index reduction of the silicon film. The paper reports that the residual between the two solvers is a single constant phase offset attributable to a fixed excess coupling length L_extra = 2.837(11) μm that remains invariant across a factor-of-two range in κ. After folding this offset into a closed-loop length correction, the agent produces a design whose FDTD-measured cross fraction is 0.498 (target 0.500).

Significance. If the reported invariance and self-consistency hold, the work illustrates that an LLM can successfully orchestrate a closed design loop between deterministic solvers in a reduced-dimensional model, reaching sub-percent accuracy on the target splitting ratio. The explicit identification of the model discrepancy as a single invariant length parameter and the resulting near-ideal performance constitute a concrete demonstration of automated photonic component design within a controlled 2D framework.

major comments (2)
  1. [Abstract] Abstract: The central claim that L_extra is invariant across the tested κ range (and therefore that the residual is a single constant phase offset) is load-bearing for the self-consistency argument, yet the manuscript provides neither the underlying residual data for the multiple designs nor the fitting procedure and statistical test used to establish invariance and the quoted uncertainty ±0.011 μm.
  2. [Abstract] The final design result (FDTD cross fraction 0.498) is obtained only after applying the L_extra correction extracted from the same set of eigenmode/FDTD pairs; without tabulated pre-correction versus post-correction FDTD outcomes or an independent validation set, the magnitude of the improvement attributable to the correction cannot be assessed.
minor comments (1)
  1. The abstract states that results are made self-consistent within the 2D effective-index model; the manuscript should explicitly state whether any 3D effects or out-of-plane leakage are neglected by construction and whether this limitation is discussed.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments and for recognizing the significance of the self-consistent design loop. We address each major comment below and will revise the manuscript to incorporate the requested supporting information.

read point-by-point responses

  1. Referee: [Abstract] Abstract: The central claim that L_extra is invariant across the tested κ range (and therefore that the residual is a single constant phase offset) is load-bearing for the self-consistency argument, yet the manuscript provides neither the underlying residual data for the multiple designs nor the fitting procedure and statistical test used to establish invariance and the quoted uncertainty ±0.011 μm.

    Authors: We agree that explicit documentation of the residual data, fitting procedure, and statistical support is necessary to substantiate the invariance claim. The current manuscript states the result but does not tabulate the per-design phase residuals or describe the linear regression used to extract L_extra and its uncertainty. In the revision we will add a supplementary table (or expanded figure) listing the eigenmode/FDTD residual phases for each tested κ, together with the regression details and the statistical criterion (e.g., goodness-of-fit or bootstrap interval) that yields the quoted ±0.011 μm uncertainty. A brief reference to this material will also be inserted in the abstract. revision: yes

  2. Referee: [Abstract] The final design result (FDTD cross fraction 0.498) is obtained only after applying the L_extra correction extracted from the same set of eigenmode/FDTD pairs; without tabulated pre-correction versus post-correction FDTD outcomes or an independent validation set, the magnitude of the improvement attributable to the correction cannot be assessed.

    Authors: We concur that the improvement due to the correction cannot be quantified without the before-and-after comparison. The manuscript reports only the final post-correction FDTD result. In the revision we will include a table (or plot) that lists the FDTD cross fractions obtained for the same geometry both before and after the L_extra length adjustment, thereby showing the magnitude of the correction. Because the correction was derived from the full ensemble of designs, an entirely independent validation set is not present; we will add an explicit statement noting this limitation and the scope of the self-consistency claim within the 2D effective-index model. revision: yes

Circularity Check

0 steps flagged

No significant circularity detected

full rationale

The paper presents an engineering workflow in which an LLM agent proposes designs, an eigenmode solver extracts κ, and an independent FDTD solver validates the response, all within a shared 2D effective-index model. The offset L_extra is characterized from residuals between the two solvers and used for length correction, but the final reported performance (FDTD cross fraction 0.498) is a direct output of the FDTD solver after the correction is applied, not a quantity that reduces to the fit by construction. The claimed invariance of L_extra across a factor-of-two range in κ supplies an independent check within the campaign. No derivation, prediction, or uniqueness claim is shown to be equivalent to its own inputs; the result is an empirical design outcome inside a self-consistent modeling framework.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The result rests on the validity of the 2D effective-index reduction and on a fitted excess length extracted from the same simulation set used for validation.

free parameters (1)
  • L_extra = 2.837(11) um
    Excess coupling length of 2.837(11) micrometers fitted from the constant phase offset between eigenmode and FDTD results.
axioms (1)
  • domain assumption The 2D effective-index reduction of the silicon film produces eigenmode and FDTD results that differ only by a single constant phase offset.
    Invoked to establish self-consistency of the design within the model.

pith-pipeline@v0.9.1-grok · 5818 in / 1493 out tokens · 36945 ms · 2026-06-26T10:07:19.695069+00:00 · methodology

0 comments
read the original abstract

We present a design agent which is a Large Language Model (LLM) that orchestrates, but does not perform, the numerical simulations to design a silicon-on-insulator (SOI) $2\times2$ directional coupler. We choose a symmetric phase-matched coupler where a lot of analytical results are available that help the design strategy. The LLM proposes candidate gap values (a geometrical dimension size) and judges convergence, while all physics is owned by deterministic solvers: a frequency-domain eigenmode solver estimates the coupling coefficient~$\kappa$ for the current design, and an independent Finite-Difference Time-Domain (FDTD) stage validates it. Both solvers operate on a common slab-projected two-dimensional (2D) effective-index reduction of the silicon film, so the design~$\kappa$ and the FDTD response are consistent by problem design; the residual between them is shown to be a single constant phase offset~$\phi$, attributable to a fixed excess coupling length $L_{\mathrm{extra}}=\SI{2.837(11)}{\micro\meter}$ that we find invariant across a factor-of-two range in~$\kappa$. Folding this offset into a closed-loop length correction, the agent delivers a $50/50$ splitter whose FDTD-measured cross fraction is $0.498$ (target $0.500$), a residual of $0.0017$. Results are made self-consistent within the 2D effective-index model; and the LLM succeeds in delivering a suitable design over a number of attempts.

Figures

Figures reproduced from arXiv: 2606.22493 by Amrit De, Md Tauhidul Islam, Saumya Biswas.

Figure 1
Figure 1. Figure 1: Agent architecture. The LLM orchestrator proposes the next gap and [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Relative variation of the predicted split plotted against refinement level. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 4
Figure 4. Figure 4: Closed-loop result. FDTD-measured cross fraction versus coupling [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Lead-in coupling length Lextra = ϕ/κ plotted against the gap. The value is more or less constant at 2.837(11) µm (dashed line, shaded ±1σ) across a factor-of-two range in κ, ascertaining that the phase offset arises from a fixed lead-in geometry and is independent of the gap. reiterate (from subsection V-A) that effective-index reductions are known to produce residual phase offsets between reduced and high… view at source ↗

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

21 extracted references · 2 canonical work pages · 1 internal anchor

  1. [1]

    Silicon photonic circuit design using rapid prototyping foundry process design kits,

    L. Chrostowski, H. Shoman, M. Hammood, H. Yun, J. Jhoja, E. Luan, S. Lin, A. Mistry, D. Witt, N. A. Jaegeret al., “Silicon photonic circuit design using rapid prototyping foundry process design kits,”IEEE Journal of Selected Topics in Quantum Electronics, vol. 25, no. 5, pp. 1–26, 2019

    2019

  2. [2]

    Silicon photonics circuit design: methods, tools and challenges,

    W. Bogaerts and L. Chrostowski, “Silicon photonics circuit design: methods, tools and challenges,”Laser & Photonics Reviews, vol. 12, no. 4, p. 1700237, 2018

    2018

  3. [3]

    Silicon photonics,

    S. Photonics, “Silicon photonics,”Topics in Applied Physics, vol. 94, 2004

    2004

  4. [4]

    Design flow automation for silicon photonics: Challenges, collaboration, and standardization,

    M. Heins, C. Cone, J. Ferguson, R. Cao, J. Pond, J. Klein, T. Korthorst, A. Bakker, R. Stoffer, M. Fierset al., “Design flow automation for silicon photonics: Challenges, collaboration, and standardization,” inSilicon Photonics III: Systems and Applications. Springer, 2016, pp. 99–156

    2016

  5. [5]

    Towards an automated design framework for large-scale photonic integrated circuits,

    S. Mingaleev, A. Richter, E. Sokolov, C. Arellano, and I. Koltchanov, “Towards an automated design framework for large-scale photonic integrated circuits,” inIntegrated Optics: Physics and Simulations II, vol. 9516. SPIE, 2015, p. 951602

    2015

  6. [6]

    Generative large language models in engineering design: opportunities and chal- lenges,

    F. Chiarello, S. Barandoni, M. M. ˇSkec, and G. Fantoni, “Generative large language models in engineering design: opportunities and chal- lenges,”Proceedings of the Design Society, vol. 4, pp. 1959–1968, 2024

    1959

  7. [7]

    A survey of research in large language models for electronic design automation,

    J. Pan, G. Zhou, C.-C. Chang, I. Jacobson, J. Hu, and Y . Chen, “A survey of research in large language models for electronic design automation,” ACM Transactions on Design Automation of Electronic Systems, vol. 30, no. 3, pp. 1–21, 2025

    2025

  8. [8]

    Toward automated simulation research workflow through llm prompt engineering design,

    Z. Liu, Y . Chai, and J. Li, “Toward automated simulation research workflow through llm prompt engineering design,”Journal of Chemical Information and Modeling, vol. 65, no. 1, pp. 114–124, 2024

    2024

  9. [9]

    ReAct: Synergizing Reasoning and Acting in Language Models

    S. Yao, J. Zhao, D. Yu, N. Du, I. Shafran, K. Narasimhan, and Y . Cao, “React: Synergizing reasoning and acting in language models,”arXiv preprint arXiv:2210.03629, 2022

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  10. [10]

    Toolformer: Language models can teach themselves to use tools,

    T. Schick, J. Dwivedi-Yu, R. Dess `ı, R. Raileanu, M. Lomeli, E. Hambro, L. Zettlemoyer, N. Cancedda, and T. Scialom, “Toolformer: Language models can teach themselves to use tools,”Advances in neural informa- tion processing systems, vol. 36, pp. 68 539–68 551, 2023

    2023

  11. [11]

    Coupled-mode theory,

    H. A. Haus and W. Huang, “Coupled-mode theory,”Proceedings of the IEEE, vol. 79, no. 10, pp. 1505–1518, 1991

    1991

  12. [12]

    Coupled-mode theory for guided-wave optics,

    A. Yariv, “Coupled-mode theory for guided-wave optics,”IEEE Journal of Quantum Electronics, vol. 9, no. 9, pp. 919–933, 1973

    1973

  13. [13]

    Coupled-mode theory for optical waveguides: an overview,

    W.-P. Huang, “Coupled-mode theory for optical waveguides: an overview,”Journal of the Optical Society of America A, vol. 11, no. 3, pp. 963–983, 1994

    1994

  14. [14]

    Liu,Photonic Devices

    J.-M. Liu,Photonic Devices. Cambridge, UK: Cambridge University Press, 2005

    2005

  15. [15]

    Okamoto,Fundamentals of Optical Waveguides, 2nd ed

    K. Okamoto,Fundamentals of Optical Waveguides, 2nd ed. Burlington, MA: Academic Press, 2006

    2006

  16. [16]

    Low-loss silicon directional coupler with arbitrary coupling ratios for broad- band wavelength operation based on bent waveguides,

    A. H. El-Saeed, A. Elshazly, H. Kobbi, R. Magdziak, G. Lepage, C. Marchese, J. Rahimi Vaskasi, S. Bipul, D. Bode, M. Ersek Filipcic, D. Velenis, M. Chakrabarti, P. De Heyn, P. Verheyen, P. Absil, F. Ferraro, Y . Ban, J. Van Campenhout, W. Bogaerts, and Q. Deng, “Low-loss silicon directional coupler with arbitrary coupling ratios for broad- band wavelength...

    work page arXiv 2024

  17. [17]

    Block-iterative frequency- domain methods for Maxwell’s equations in a planewave basis,

    S. G. Johnson and J. D. Joannopoulos, “Block-iterative frequency- domain methods for Maxwell’s equations in a planewave basis,”Optics Express, vol. 8, no. 3, pp. 173–190, 2001

    2001

  18. [18]

    MEEP: A flexible free-software package for electromagnetic simulations by the FDTD method,

    A. F. Oskooi, D. Roundy, M. Ibanescu, P. Bermel, J. D. Joannopoulos, and S. G. Johnson, “MEEP: A flexible free-software package for electromagnetic simulations by the FDTD method,”Computer Physics Communications, vol. 181, no. 3, pp. 687–702, 2010

    2010

  19. [19]

    Integrated circuits for the millimeter through optical frequency range,

    R. Knox and P. Toulios, “Integrated circuits for the millimeter through optical frequency range,” inProc. Symp. Submillimeter Waves, vol. 20. Brooklyn, NY , 1970, pp. 497–515

    1970

  20. [20]

    Inverse-designed low-index-contrast structures on a silicon photonics platform for vector–matrix multiplication,

    V . Nikkhah, A. Pirmoradi, F. Ashtiani, B. Edwards, F. Aflatouni, and N. Engheta, “Inverse-designed low-index-contrast structures on a silicon photonics platform for vector–matrix multiplication,”Nature Photonics, vol. 18, no. 5, pp. 501–508, 2024

    2024

  21. [21]

    Yariv and P

    A. Yariv and P. Yeh,Photonics: Optical Electronics in Modern Com- munications, 6th ed. New York: Oxford University Press, 2007

    2007