3D surface profiling via photonic integrated geometric sensor

Chunhui Yao; Huiyu Huang; Liang Ming; Minjia Chen; Qixiang Cheng; Richard Penty; Rui Ma; Ting Yan; Wanlu Zhang; Xin Du

REVIEW 3 major objections 6 minor 19 references

3D surface profiling via photonic integrated geometric sensor

T0 review · 3 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Photonic chip maps 3D surface shape from reflected light alone

desk verdict A believable proof-of-concept for a compact PIC-based surface profiler, but the flagship generalization claim (the lion emblem) is only qualitatively supported and the training labels are underdescribed. read the letter →

arxiv 2506.23187 v1 pith:J27YR5PE submitted 2025-06-29 physics.optics

Ziyao Zhang , Yizhi Wang , Chunhui Yao , Huiyu Huang , Rui Ma , Xin Du , Wanlu Zhang , Zhitian Shi

show 6 more authors

Minjia Chen Ting Yan Liang Ming Yuxiao Ye Richard Penty Qixiang Cheng

This is my paper · ORCID

classification physics.optics

keywords photonicintegratedcircuitssurfaceprofilometrymicroringresonatorarrayspectralencoding1Dconvolutionalneuralnetwork3Dprintedmicrostructuresgeometricsensingopticalmetrology

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

This paper tries to show that microscale 3D surface profiling can be reduced to a single reflected-light measurement handled by a compact photonic chip and a neural network, replacing bulky interferometric benches. The authors built a silicon nitride circuit whose six thermo-optically tuned microring resonators create 729 pseudo-random spectral sampling states; the intensity recorded at each state forms a fingerprint of the light reflected from the sample. A lightweight one-dimensional convolutional network decodes that fingerprint directly into a thickness value, trained only on data from a smoothly varying 3D-printed slope. On a separately printed slope the sensor predicted thickness with an average deviation of 18.5 micrometres, and on an unseen lion-shaped emblem it resolved two height levels with features as small as 30 micrometres. If the claim holds, surface metrology could move from laboratory benches to handheld, real-time, in-situ scanners.

What carries the argument

The load-bearing object is the programmable spectral sampler: six over-coupled microring resonators with distinct free spectral ranges, cascaded and individually thermo-optically adjusted among three states to form 729 pseudo-random transmission masks. It acts as a physical random projection that compresses the reflected spectrum into an information-rich intensity vector. The 1D-CNN decoder is matched to this structure: its width-3 kernels exploit the fact that masks differing in only one ring's setting produce adjacent, physically related channels, while the fully connected head reduces the projected features to a single thickness output.

What would settle it

Scan flat regions of equal thickness but different surface roughness, tilt, or focus; if predicted thickness shifts with those changes, the thickness-only spectral mapping is false.

Watch

Extended reading notes

Core claim

The central claim is that surface height can be recovered from a learned mapping between the sample's reflected spectrum and its local thickness, so no interferometric reference arm, fringe analysis, or nonlinear fitting is needed. The photonic encoder physically projects the reflected light through a pseudo-random transmission matrix—six over-coupled microrings, each switched among three phase states—yielding 729 distinct intensity readings per scan point. The 1D-CNN then regresses these readings to thickness, with a small convolution kernel because neighboring channels share physical ring settings and therefore carry related information. Because training used only the slope sample, the successful reconstruction of the unrelated lion emblem is the paper's evidence that the encoder–decoder pair learns the underlying spectral-to-thickness relation rather than memorizing the training surface.

Load-bearing premise

The reflected spectrum at each scan point depends only on local sample thickness, so a decoder trained on one smooth slope generalizes to arbitrary lateral patterns without retraining.

Editorial extensions

If this is right

Surface profiling can be performed with a 0.4 mm² chip and a single photodetector, with no reference arm or interferometric alignment.
The detection-to-prediction pipeline collapses to one neural-network inference pass, so throughput is set by modulation and scanning speed rather than by reconstruction algorithms.
A decoder trained once on a smooth slope transfers to unseen sharply stepped patterns, suggesting the learned mapping is tied to sample thickness rather than to a specific surface.
Higher lateral resolution and faster scanning should follow from better galvo control, smaller focal spots, and faster phase modulation, without changing the core encoder-decoder design.
Integrating a light source and detector on the same chip could turn the profiler into a standalone, portable metrology module.

Reading between the lines

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

The same encoder-decoder pipeline should extend to other scalar measurands encoded in reflection, such as thin-film thickness, refractive index, or surface roughness, because the network never sees the physical quantity, only intensity fingerprints.
The robustness claim would be tested harder by printing samples that vary thickness and roughness independently; if the network's thickness output shifts when roughness changes, the current mapping is not purely thickness-based.
Because the 729 masks are programmable, the same chip could be repurposed for different sample classes by retraining the decoder on a new calibration set, turning one hardware design into a multi-task geometric sensor.
A systematic study of how many masks are needed for a given accuracy could give a quantitative trade-off between scan time and fidelity; the paper uses 729 as a balanced choice.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

Bottom line: this is a believable proof-of-concept, not yet a demonstrated profiler. The genuinely new bit is taking the programmable MRR spectral sampler this group has used for reconstructive spectrometers and pairing it with a 1D-CNN that regresses surface thickness directly from 729 photodetector readings. That is a real extension, and the hardware is compact (0.4 mm² SiN chip) and alignment-free. What is done well: the training data are substantial (90k points on a 20–200 µm slope), the model is tested on a separately printed slope sample scanned over a 2D area, and the reported 17.5 µm test RMSE and 18.5 µm average deviation are consistent with no obvious overfitting. That is a solid reproducibility check for the smooth-slope case.

The soft spots are real, and the stress-test note lands. The lion emblem is the paper’s own evidence for generalization to arbitrary patterns, but the evaluation is entirely qualitative: the two thickness levels are visually resolved, and the right eye is described as spanning "approximately three pixels (30 μm)." There is no quantitative comparison to the CAD-defined 40 µm and 150 µm plateaus, no error map, no confusion metric. With what is reported, you cannot distinguish genuine thickness profiling from lateral pattern separation. The ground-truth procedure is also underdescribed: "standard benchtop spectrometer" is not a method specification, and if the thickness labels come from a spectral-reflectance inversion similar to the sensor’s own physical principle, the evaluation is more circular than the text admits. There is no baseline comparison against even a simple spectral fit, and no code or data release. Minor, but worth saying: the abstract describes the result as "high-fidelity," while 18.5 µm average error on a 200 µm range is roughly 10% of full scale.

I do not think these flaws sink the core idea. They do mean the central generalization claim is currently underdetermined. A serious referee could fix this with one quantitative experiment on the lion (or on a few discrete-height patterns) plus a clear description of the ground-truth method. This paper deserves peer review rather than desk rejection — it is a genuine attempt to build something hard — but it should come back with revisions. For readers in integrated photonics and compact optical metrology, it is worth a look as a proof-of-concept; just do not cite it as a validated metrology technique yet.

Referee Report

3 major / 6 minor

Summary. The manuscript proposes an ultracompact surface profiler built from a silicon nitride photonic integrated circuit containing six cascaded microring resonators. The circuit operates as a programmable pseudo-random spectral encoder that generates 729 sampling states, each producing a single photodetector reading from light reflected by the sample. A one-dimensional convolutional neural network (1D-CNN) decodes the 729-dimensional vector into a local thickness value. The network is trained on more than 90,000 measurements from a 3D-printed smooth slope with thickness varying from 20 to 200 μm, and the authors report a test RMSE of about 17.5 μm and an average deviation of 18.5 μm on a separately printed slope. They further demonstrate the approach on a 3D-printed lion emblem with two nominal thickness levels, 40 μm and 150 μm, claiming that the two levels are clearly resolved and that features as small as 30 μm are captured. The central claim is that this device enables high-fidelity, fast-scanning-rate thickness identification for both smooth and intricate patterned surfaces without interferometric setups or nonlinear fitting.

Significance. If the claims are substantiated, this work would represent a useful advance in compact, alignment-free optical metrology, combining programmable photonic encoding with a lightweight neural decoder. The demonstrated hardware is small (0.4 mm² chip) and the inference pipeline bypasses conventional spectral reconstruction, which could be attractive for in-situ or handheld profiling. The paper also provides a reasonable physical rationale for using a 1D-CNN on the sequentially configured sampling channels. However, the current evidence is insufficient to fully establish the headline claim of high-fidelity identification of intricate patterns: the only quantitative validation is on a smooth slope, and the complex-pattern test is reported qualitatively. The ground-truth reference method is also underspecified. These issues are load-bearing for the paper's central claim, so the result is plausible and worth pursuing but not yet fully demonstrated.

major comments (3)

[Experimental Validation and Analysis (Fig. 5)] The lion emblem demonstration is evaluated only qualitatively. The text states that 'the two thickness levels are clearly resolved' and that the right eye spans roughly three pixels (30 μm), but no quantitative comparison is made between the predicted thickness map and the CAD-defined plateaus of 40 μm and 150 μm. This is load-bearing because the claim of 'high-fidelity thickness identification for intricate 3D printed emblem structures' (Abstract) requires demonstrating that the decoder outputs absolute thickness values, not merely that it separates two intensity classes. Please provide per-level error statistics (e.g., mean absolute error and RMSE for the 40 μm and 150 μm regions separately), an error map, and a quantitative measure of lateral edge resolution, such as a line profile across a known step edge.
[Data Acquisition and Training Strategy] The ground-truth thickness labels are described as 'retrieved' with a 'standard benchtop spectrometer,' but the specific instrument, measurement principle, calibration procedure, and uncertainty are not given. If the benchtop spectrometer infers thickness from the same spectral reflectance signal that the PIC encodes, the evaluation may be partly circular. Please specify the reference method (e.g., spectral reflectance fitting, white-light interferometry, or confocal profilometry), report its independently known accuracy, and, ideally, validate the ground truth on a calibrated step-height standard. Without this, the reported 18.5 μm average deviation cannot be properly interpreted.
[Experimental Validation and Analysis (slope generalization)] The only quantitative test of the trained network is on a second slope sample with the same nominal geometry as the training data, differing only in fabrication batch and scan area. This does not rule out the possibility that the decoder relies on confounded cues that correlate with height on the slope but not in arbitrary lateral patterns, such as local reflectivity, defocus, or edge scattering. The manuscript's implicit assumption that the 729-channel vector is a deterministic function of local thickness alone is central to the profiler claim, but it is not directly tested. Please add a quantitative transfer experiment to surfaces with lateral patterns different from the training distribution—the lion scan, once quantified, would serve this purpose—or provide a direct physical test (e.g., repeated measurements under varying focus, tilts, or surface roughness) showing the encoder response is dominated by thickness.

minor comments (6)

[Experimental Validation and Analysis] There is a typo: 'The predicated 3D profile' should read 'The predicted 3D profile.'
[Materials and Methods (Fabrication)] The phrase 'is used here for s the balanced printing speed and quality' contains a stray 's'; please correct it.
[Results, Design of Photonic Encoder] The text refers to 'Fig.2 (d) for combined transmission matrix,' but Figure 2 appears to contain only panels (a) and (b); please either add the referenced panel or correct the citation.
[Abstract and Conclusion] The paper claims 'fast-scanning-rate thickness identification,' but no scan rate, measurement time per point, or total acquisition time is reported anywhere. Please provide at least a per-point inference time and a scan throughput estimate.
[Results, Surface Profiling Decoder Based on 1D-CNN] The physical rationale for channel ordering is plausible, but it is not empirically tested. An ablation that permutes the channel order or compares the 1D-CNN against an MLP with the same number of parameters would strengthen the architectural justification.
[Data Acquisition and Training Strategy] The training and test RMSE values (8.2 μm vs 17.5 μm) are reported as single numbers without confidence intervals or repeated runs. Reporting the variance across multiple training runs or data splits would help assess stability.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the sensor claim is validated on held-out samples via an empirical encoder-decoder, with no self-referential derivation.

full rationale

The paper makes no analytic derivation claim; the device encodes reflected spectra through a PIC and decodes thickness with a 1D-CNN trained on a 90,000-point slope dataset and evaluated on held-out slope and lion samples. The training labels are stated to come from "a standard benchtop spectrometer" (Data Acquisition and Training Strategy), not from the PIC output itself, so the target is not defined by the network input on the evidence given. The CNN architecture is presented as a physical hypothesis ("we hypothesize that the CNN part of our model can project the sample PD readings into higher dimensional spaces"), not as a result derived from its own assumptions. Self-citations (refs 10, 14-16) are background on integrated spectrometers and do not carry the load of the profiling claim. The lion-emblem validation is qualitative and lacks a quantitative thickness error map against the CAD model, which is a completeness/correctness concern but not circularity. No equation in the paper reduces a prediction to its fitted input, and no uniqueness or ansatz is imported from prior work by the same authors. Therefore no circular step can be exhibited from the manuscript text.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The paper introduces no new physical entities. It relies on the trained neural network weights as free parameters, and on several domain assumptions about the relationship between reflectance and thickness, the accuracy of the ground truth, the stability of the sampler, and the relevance of channel ordering. The most significant assumption is that the reflectance spectrum is a deterministic function of thickness alone, which underpins the generalization claim.

free parameters (3)

Neural network weights = Not reported
All CNN and MLP weights are fitted to the 90,000 training examples via Adam optimization. The central claim of thickness identification depends entirely on these fitted weights, which are not released.
Number of sampling states (729) = 729 (3^6)
The choice of three discrete phase states per ring and six rings is hand-selected as a trade-off between encoding performance, measurement time, and computational cost, as stated in the Results section.
Hyperparameters (learning rate, decay, kernel size, architecture) = lr=1e-4, decay=1e-5, kernel=3, layers c1-c3 and m1-m3
These are chosen by the authors and are not derived from data or theory; they affect the network's ability to learn the mapping from PD readings to thickness.

assumptions (4)

domain assumption The reflectance spectrum of the sample encodes its local thickness.
The entire sensing concept relies on a stable mapping between the reflected spectrum and surface height. The physical mechanism (e.g., thin-film interference or scattering) is not explicitly explained in the paper, and this is invoked implicitly in the Introduction and Data Acquisition.
domain assumption Ground truth thickness values from the benchtop spectrometer are accurate and independent.
The paper states 'we used standard benchtop spectrometer to retrieve the target thickness of the structure' without describing the retrieval method or its calibration. If this ground truth is derived from the same spectral reflectance physics, the evaluation becomes circular.
domain assumption The 729 sampling states are reproducible and stable over time.
The phase states are generated by a pre-calibrated voltage look-up table and rely on active TEC temperature control. The paper does not report repeatability measurements across multiple scans or days, so the stability of the encoder is assumed.
ad hoc to paper Channel ordering by sequential ring tuning carries spatial relevance that a 1D-CNN can exploit.
The authors argue that masks differing in one ring are physically more related, justifying a small kernel width of 3. This is a modeling heuristic specific to this design and is not rigorously derived.

how reviews work

0 comments

Cite this review

Pith. "Pith review of 3D surface profiling via photonic integrated geometric sensor." pith.science (2026). https://pith.science/paper/J27YR5PE

@misc{pith2026250623187,
  author       = {Pith},
  title        = {Pith review of: 3D surface profiling via photonic integrated geometric sensor},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/J27YR5PE}},
  note         = {Machine review of arXiv:2506.23187}
}

read the original abstract

Measurements of microscale surface patterns are essential for process and quality control in industries across semiconductors, micro-machining, and biomedicines. However, the development of miniaturized and intelligent profiling systems remains a longstanding challenge, primarily due to the complexity and bulkiness of existing benchtop systems required to scan large-area samples. A real-time, in-situ, and fast detection alternative is therefore highly desirable for predicting surface topography on the fly. In this paper, we present an ultracompact geometric profiler based on photonic integrated circuits, which directly encodes the optical reflectance of the sample and decodes it with a neural network. This platform is free of complex interferometric configurations and avoids time-consuming nonlinear fitting algorithms. We show that a silicon programmable circuit can generate pseudo-random kernels to project input data into higher dimensions, enabling efficient feature extraction via a lightweight one-dimensional convolutional neural network. Our device is capable of high-fidelity, fast-scanning-rate thickness identification for both smoothly varying samples and intricate 3D printed emblem structures, paving the way for a new class of compact geometric sensors.

Figures

Figures reproduced from arXiv: 2506.23187 by the authors.

**Figure 1.** Schematic illustrations of (a) conventional benchtop interferometric spectrometry systems and (b) our integrated on-chip geometric sensing scheme, respectively [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

19 extracted references · 19 canonical work pages

[1]

Leach, Ed., Optical Measurement of Surface Topography (Springer, Berlin, Heidelberg, 2011; http://link.springer.com/10.1007/978-3-642-12012-1)

R. Leach, Ed., Optical Measurement of Surface Topography (Springer, Berlin, Heidelberg, 2011; http://link.springer.com/10.1007/978-3-642-12012-1)

work page doi:10.1007/978-3-642-12012-1 2011
[2]

K. J. Stout, L. Blunt, Three Dimensional Surface Topography (Elsevier, 2000)

work page 2000
[3]

Y. Zhou, D. Grass, W. S. Warren, M. C. Fischer, Non-destructive three-dimensional imaging of artificially degraded CdS paints by pump-probe microscopy. J. Phys. Photonics 6, 025013 (2024)

work page 2024
[4]

P. J. de Groot, A review of selected topics in interferometric optical metrology. Rep. Prog. Phys. 82, 056101 (2019)

work page 2019
[5]

S. Feng, L. Zhang, C. Zuo, T. Tao, Q. Chen, G. Gu, High dynamic range 3D measurements with fringe projection profilometry: a review. Meas. Sci. Technol. 29, 122001 (2018)

work page 2018
[6]

Shimizu, L.-C

Y. Shimizu, L.-C. Chen, D. W. Kim, X. Chen, X. Li, H. Matsukuma, An insight into optical metrology in manufacturing. Meas. Sci. Technol. 32, 042003 (2021)

work page 2021
[7]

LeCun, Y

Y. LeCun, Y. Bengio, G. Hinton, Deep learning. Nature 521, 436–444 (2015)

work page 2015
[8]

G. V. der Sande, D. Brunner, M. C. Soriano, Advances in photonic reservoir computing. Nanophotonics 6, 561–576 (2017)

work page 2017

Show all 19 references

[9]

S. Yuan, C. Ma, E. Fetaya, T. Mueller, D. Naveh, F. Zhang, F. Xia, Geometric deep optical sensing. Science 379, eade1220 (2023)

2023
[10]

A. Li, C. Yao, J. Xia, H. Wang, Q. Cheng, R. Penty, Y. Fainman, S. Pan, Advances in cost -effective integrated spectrometers. Light Sci Appl 11, 174 (2022)

2022
[11]

M. A. Butt, X. Mateos, R. Piramidowicz, Photonics sensors: A perspective on current advancements, emerging challenges, and potential solutions (Invited). Physics Letters A 516, 129633 (2024)

2024
[12]

A. Bag, M. Neugebauer, U. Mick, S. Christiansen, S. A. Schulz, P. Banzer, Towards fully integrated photonic displacement sensors. Nat Commun 11, 2915 (2020)

2020
[13]

Z. Zhi, Q. Na, Q. Xie, B. Chen, Y. Li, X. Liu, X. Li, L. Wang, G. Lo, J. Song, On -chip generation of Bessel– Gaussian beam via concentrically distributed grating arrays for long-range sensing. Light Sci Appl 12, 92 (2023)

2023
[14]

C. Yao, K. Xu, W. Zhang, M. Chen, Q. Cheng, R. Penty, Integrated reconstructive spectrometer with programmable photonic circuits. Nat Commun 14, 6376 (2023)

2023
[15]

C. Yao, J. Ma, N. Wang, P. Bao, W. Zhuo, T. Zhang, W. Zhang, K. Xu, T. Yan, L. Ming, Y. Ye, T. Hasan, I. White, R. Penty, Q. Cheng, Optical Convolutional Spectrometer. arXiv arXiv:2502.08749 [Preprint] (2025). https://doi.org/10.48550/arXiv.2502.08749

2025 doi arXiv
[16]

C. Yao, K. Xu, T. Lin, J. Ma, C. Yao, P. Bao, Z. Shi, R. Penty, Q. Cheng, Benchmarking Reconstructive Spectrometer with Multiresonant Cavities. ACS Photonics 11, 3730–3740 (2024)

2024
[17]

T. Wang, M. M. Sohoni, L. G. Wright, M. M. Stein, S.-Y. Ma, T. Onodera, M. G. Anderson, P. L. McMahon, Image sensing with multilayer nonlinear optical neural networks. Nat. Photon. 17, 408–415 (2023)

2023
[18]

Mennel, J

L. Mennel, J. Symonowicz, S. Wachter, D. K. Polyushkin, A. J. Molina-Mendoza, T. Mueller, Ultrafast machine vision with 2D material neural network image sensors. Nature 579, 62–66 (2020)

2020
[19]

S. Lee, R. Peng, C. Wu, M. Li, Programmable black phosphorus image sensor for broadband optoelectronic edge computing. Nat Commun 13, 1485 (2022)

2022

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [1]

Leach, Ed., Optical Measurement of Surface Topography (Springer, Berlin, Heidelberg, 2011; http://link.springer.com/10.1007/978-3-642-12012-1)

R. Leach, Ed., Optical Measurement of Surface Topography (Springer, Berlin, Heidelberg, 2011; http://link.springer.com/10.1007/978-3-642-12012-1)

work page doi:10.1007/978-3-642-12012-1 2011

[2] [2]

K. J. Stout, L. Blunt, Three Dimensional Surface Topography (Elsevier, 2000)

work page 2000

[3] [3]

Y. Zhou, D. Grass, W. S. Warren, M. C. Fischer, Non-destructive three-dimensional imaging of artificially degraded CdS paints by pump-probe microscopy. J. Phys. Photonics 6, 025013 (2024)

work page 2024

[4] [4]

P. J. de Groot, A review of selected topics in interferometric optical metrology. Rep. Prog. Phys. 82, 056101 (2019)

work page 2019

[5] [5]

S. Feng, L. Zhang, C. Zuo, T. Tao, Q. Chen, G. Gu, High dynamic range 3D measurements with fringe projection profilometry: a review. Meas. Sci. Technol. 29, 122001 (2018)

work page 2018

[6] [6]

Shimizu, L.-C

Y. Shimizu, L.-C. Chen, D. W. Kim, X. Chen, X. Li, H. Matsukuma, An insight into optical metrology in manufacturing. Meas. Sci. Technol. 32, 042003 (2021)

work page 2021

[7] [7]

LeCun, Y

Y. LeCun, Y. Bengio, G. Hinton, Deep learning. Nature 521, 436–444 (2015)

work page 2015

[8] [8]

G. V. der Sande, D. Brunner, M. C. Soriano, Advances in photonic reservoir computing. Nanophotonics 6, 561–576 (2017)

work page 2017