Ultra-Compact Inverse-Designed Integrated Photonic Matrix Compute Core

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

The manuscript proposes an ultra-compact inverse-designed photonic matrix compute core (PMCC) based on the GLINT algorithm, which demonstrates great potential for improving the integration density of photonic neural networks. To further enhance the rigor and clarity of the paper, I would like to raise the following questions and suggestions:

1.The manuscript reports a recognition accuracy of 99.05% on the MNIST dataset, but there is no comparison with conventional electronic neural networks. To what extent does this result indicate performance degradation compared to traditional neural networks?

2.In line 90, the authors mentioned the number of phase shifters, which seems to be 4n instead of the stated 2n. The authors are advised to verify this point.

3.For the proposed compact phase shifter, what is the length (L) of the tunable-width section? In the design, how is the transition implemented from the fixed-width waveguide (0.5 μm) to the maximum tunable width (0.5–1.0 μm)?

4.In lines 207 and 219, the text states that the width W varies between 0.5 μm and 1.0 μm. However, in Figure 3(b) and (c), the data for W = 0.9–1.0 μm are missing. Please ensure consistency between the text and the figures.

5.In line 236, a phase deviation of π/20 is said to correspond to a 10% average relative phase error (RPE). Relative to which phase shift is this percentage calculated, and how is it derived?

6.The manufacturing error model for the power divider qualitatively describes the characteristics of removing "islands" and merging, but does not provide a quantitative description. What is the specific critical feature size?

7.Is the reported recognition accuracy accuracy (99.05%) obtained through the electromagnetic simulation of the whole cascade or calculated through the transmission matrix of a single component of the cascade?

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

Dear Editor,

The authors have submitted a paper titled " ultra compact inverse-designed integrated photonic matrix compute core". The manuscript presents major shortcomings that the authors need to address. Therefore, I do not recommend the manuscript for publication in its current form.

 

1- Remove abbreviations like GLINT and SOI from the abstract, defining them in the main text instead.

 

2- The provided statement on page 1, lines 31-33, is a general claim about the benefits of optical neural networks over traditional artificial neural networks. Therefore, citing five separate references for this single, common assertion is unnecessary. A single, representative reference is sufficient to support this claim.

 

3- To help readers get a full understanding of the manuscript and the research stages, introduce its different parts at the end of the introduction.

 

4- What is the innovation of your study?

 

5- Describe in detail the software used for simulation and the applied boundary conditions.

 

6- Have you considered the nonlinear effects of silicon during simulations and its impact on the performance of the structure?

 

7- Provide the transmission loss for your proposed structure.

 

8- An important issue in this structure is crosstalk. It is necessary to include its formula in the text, report the amount of crosstalk, and discuss it.

 

9- The impact of environmental factors (such as temperature, air pressure, and humidity) on device performance and results should be discussed.

 

10- In order to evaluate the results obtained from the simulations, it is necessary to compare the important results obtained from the proposed structure and previous similar research in a table.

 

11- On what basis were the values of structural parameters selected? Can their optimality be asserted?

 

Kind regards

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

The manuscript presents an inverse-designed photonic matrix compute core (PMCC) on an SOI platform, achieving remarkable compactness, high integration density (~26,000 units/mm²), and strong performance in ONN applications. The topic is timely and relevant, and the work demonstrates technical novelty through the GLINT algorithm for designing compact symmetric power splitters and phase shifters. The manuscript is generally well-structured and clearly written, though some clarifications and additional discussions are needed.

Question 1. In the introduction, comparisons with state-of-the-art ONNs could be expanded with more quantitative benchmarking beyond density. How does accuracy compare to prior compact architectures (e.g., DONNs or PCM-based ONNs)?

Question 2. In Introduction to the principles of the PMCC section, Could the authors include a schematic example of signal flow for a smaller (e.g., 4×4) case to illustrate the connectivity rules? It would be helpful to explain how scalability (increasing n) impacts optical loss, training complexity, and power consumption.

Question 3. In Section 3, Compact Symmetric Power Splitter, have the authors tested wavelength dependence of the splitter performance? Figure 2(e) shows transmission, but the bandwidth tolerance should be quantified.

Question 4. In section 4 (Compact Phase Shifter), The reported phase error tolerance (±7% under ±50 nm) is encouraging, but could the authors comment on thermal stability in practical environments?

Question 5. In Section 5 (ONN Demonstration), The training methodology (gradient-based phase tuning) is well explained, but details on convergence speed, computational resources, and comparison with electronic training should be included.

Question 6. In Section 6 (Fabrication-Error Simulation), Could the authors clarify whether the >80% accuracy robustness metric corresponds to absolute recognition accuracy or relative retention of baseline performance?

Minor Comments

• Please define all acronyms (e.g., OLC) at first mention in the main text, not only in figure captions.
• Figures 2–4 would benefit from higher contrast and labeling clarity for readability.
• The conclusion section could emphasize potential application areas (e.g., real-time inference, edge AI) more explicitly.

Recommendation:

The manuscript presents significant contributions to inverse-designed photonic computing. After addressing above mentioned comment, it will be suitable for publication.

Comments for author File: Comments.pdf

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 2 Report

Comments and Suggestions for Authors

Dear Editor,

Based on the raised comments, the manuscript has been well polished, so I recommend it for publication.

Kind regards