Global–Local Cooperative Optimization in Photonic Inverse Design Algorithms

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

The authors propose a Global-Local Integrated Topology (GLINT) inverse design algorithm and demonstrate its effectiveness through the design and fabrication of a waveguide bending and wavelength division multiplexers. There are several technical questions and suggestions for clarification:

1. In the local optimization stage, circular elements of smaller sizes are flipped randomly. could the authors indicate whether this process generates structures that may present manufacturing challenges, such as small holes or isolated islands. The manuscript would benefit from discussion of manufacturing constraints.
2. The authors employed circular rather than square unit cells during the optimization process. Is there any difference between the two in terms of optimization efficiency and structural integrity (e.g., occurrence of discontinuous patterns or non-manufacturable features)?

3.For devices particularly sensitive to structural morphology (e.g., wavelength multiplexers), could the randomly generated circular features potentially lead to performance degradation after fabrication. The analysis of manufacturing tolerance would enhance the manuscript's persuasiveness.

4. For functionally complex structures, it is difficult to find suitable initial structure, many algorithms (e.g., DBS and adjoint method) exhibit dependence on initial structures. Does proposed method in the manuscript exhibit similar issue?
5. Both topology optimization based on adjoint method and DBS method can achieve similar optimization outcomes. Could the authors provide a comparative analysis between the proposed method and these two approaches in terms of optimization efficiency and device performance?

6.What are the convergence criteria for the optimization? Are maximum iterations or performance thresholds defined?

7. Since the method relies on random flipping of unit structures, how stable are the device performance results over multiple optimization runs?

Comments for author File: Comments.pdf

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

Comments:

1. The authors should clearly mention that the inverse design modeling was done in 3D FDTD. This way the long simulation times are justified and readers who would read the paper within a limited time after seeing the long optimization time would still continue reading.
2. The biggest advantage I can see from the proposed inverse design method is not needing to use binarization. This aspect should be mentioned in the abstract. 
3. While the CPU name was mentioned. I wonder how much of it was actually utilized. All 20 cores? For example if the method would be run with Lumerical FDTD using enterprise licence of the 20 possible cores only 4 would be utilized. Whereby it is understandable that the simulation time is long.
4. In Fig.5(d) of the SEM image it is not fully understood if all of the bends are with using the inverse designed bend. On this topic if possible the bend loss by bend count could be also mentioned.
5. Regarding line 258. Is it possible to explain a bit more clearly why the configuration lies at the local optimum?
6. Running 1000+ iterations while the result might yield a good pattern seems a bit too much to reach the final pattern. Especially considering that each iteration was run in 3D FDTD. - Justifying this is hard but necessary. 
7. A comparison or benchmark table or value should be added to the manuscript if possible.
8. Is it possible to add a post fabrication SEM topology to the FDTD simulator? This way the authors could correlate the results to measurements better. (Not necessarily)

Questions:

1. Does this GLINT inverse design method work if a base layout design is used? If yes this should be mentioned a bit more clearly. On the same topic. Would it also work if the initial model is empty space?
2. Is it possible to combine 2D initial optimization and then 3D optimization to reduce the time?
3. Why was the WDM structure channels spread out so much? Spaced by 100nm. Was it to reduce optimization time? For example if it is done with topology inverse design the further away the channels the shorter optimization time.

 

 

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

The paper introduces the GLINT (Global-Local Integrated Topology) algorithm as a novel optimization framework for the inverse design of photonic devices. The authors propose a dual-phase iterative strategy combining global search and local refinement, circumventing key limitations of both density topology optimization (DTO) and direct binary search (DBS). The GLINT algorithm achieves high design resolution (20 nm × 20 nm) and efficient convergence. The effectiveness of the method is demonstrated through the design and experimental validation of three devices: A dual-port WDM, A 90-degree bending waveguide, and A three-port WDM. The reported results indicate improvements in device compactness, design freedom, and computational efficiency. Here are some areas for improvement:

1. While the GLINT framework is positioned as a general optimization platform, the current scope is restricted to a specific class of passive photonic components.
2. The algorithm lacks comparative benchmarking against state-of-the-art methods beyond general commentary. Include quantitative comparisons with state-of-the-art methods like LST-DTO, rotatable DBS, or adjoint methods on similar devices.
3. The reproducibility of the algorithm (e.g., access to code, parameters used, and simulation settings) is not fully addressed. For a computational algorithm, at least pseudocode would aid reproducibility.
4. Discuss how sensitive the algorithm is to initial values of Gth, Lth, RB, and RS.
5. The GLINT algorithm currently focuses on binary material structures (silicon/silicon dioxide). Extension to grayscale or multi-material structures is not explored.
6. Lack of scalability demonstration for large-scale or active devices.
7. Simulation platform details (e.g., FDTD/FEM solvers, grid resolution, boundary conditions) are not exhaustively detailed.
8. Further elaborate on spatial filtering or simplification techniques mentioned for improving fabrication fidelity.
9. The abstract could better emphasize the technical novelty of GLINT in contrast to prior methods.
10. The acronym GLINT should be clearly defined at its first mention, not mid-sentence.
11. Tables and figure legends could benefit from enhanced formatting to aid readability.

 

The manuscript presents a technically sound, practically validated, and innovative optimization framework. While it is not suitable for publication in its current form, but major revisions are recommended to improve reproducibility, benchmarking, and clarity. The contributions are significant, and with slight refinements, this paper will have strong impact in the field of computational photonics

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 4 Report

Comments and Suggestions for Authors

To improve the quality and clarity of the manuscript, I recommend that the authors consider the following revisions:

1. Abstract: Several abbreviations are used without prior definition. The authors are encouraged to introduce each abbreviation upon first use to ensure clarity for readers unfamiliar with the terminology.

2. Equation (1): It is advisable to first present the general expression for the Figure
   of Merit (FOM). Following this, a specific example or context of its application may be included to enhance comprehension.

3. Figure 1: The caption and/or accompanying discussion should be expanded to provide a more complete description of the figure’s content and its relevance to the proposed method.

4. Figures 2 and 3: The descriptions of these figures should be elaborated, with particular attention to the subfigures (a, b, c, and d). Each subfigure should be clearly described to help the reader interpret the presented results.

5. Normalized Transmission Ratio: The corresponding equation should be introduced and explained in the main body of the text before it is referenced or used in the figure captions.

6. Algorithm Performance: Additional information is needed regarding the computational efficiency of the proposed algorithm. Specifically, how fast does it operate, and are there any constraints related to substantial miniaturization of the photonic structures?

7. Signal Integrity: The manuscript would benefit from a discussion on potential signal losses. Additionally, does the signal-to-noise ratio (SNR) vary under different conditions or configurations?

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 2 Report

Comments and Suggestions for Authors

Comments:
1.In the SEM image I see only 8 bending regions not 12. Is it because the whole pattern image where we do have 12 bending regions image was taken zoomed in including only 8 bends? Please correct as needed.

2. Do add this comment also if my understanding is correct (with the author’s wordings of course). In the defined add/remove pixel is as 20 nm x 20 nm. With this the amount of steps is as you have included in the optimization. If however it has been increased to 40 nm x 40 nm the area of adding is reduced by 4x thus the iteration numbers would also lowered at the cost of having lower FOM. How much depends on the users target design.

3. I understand that the main part of the manuscript is about the new optimization method. However slightly more detailed information regarding device fabrication is also needed. Current explanation seems to be a bit too vague.
For example (Author’s should modify it as needed): The device was fabricated using industrial standard approach of EBL exposure (with xxx resist at xxx kv and xxx used for development), followed by dry etching (with ICP-RIE or RIE; using XXX gas) and wet etching cleaning (with xxx) to remove resist. With pattern precision small as xxx nm.
Another approach if it was manufactured by MWP the authors can refer readers to it. But even this way the manufacturing precision of possible should also be mentioned.

Suggestions for future:

1.To get more comparable experiment and simulation results I would suggest to take also into account the side angle in with the code if possible. For example taking the design layer to be exactly from the center and then the top -half_sidewall_angle_difference and bottom as opposite to this. This approach has proven to be suitable when inverse designing lithium niobate structures.

2.Another aspect is to apply at the end for you pattern a post fabrication with fabrication aware modified design. A guide to utilize this is published online by tidy3d as an example (https://www.flexcompute.com/tidy3d/examples/notebooks/AdjointPlugin14PreFab/)

3. For a better result FOM could be a combination of transmission and reflection by a certain ratio. This has helped me to even do initial optimizations with resonant like designs and reduce the simulation time (by reducing the required auto-shutoff level).

 

 

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

The authors have shown exemplary diligence in their revisions. The manuscript is now a concise, well-supported, and valuable contribution to the field.

The paper is ready for publication.

Author Response

We sincerely thank the reviewers for their further review of our revised manuscript.