{"id":"fe291a3d-aedb-4231-a355-1c481f9a27ee","arxiv_id":"1909.11136","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A regularized inverse-designed silicon metamaterial splits 1520 nm and 1580 nm light from a 10-um-wide waveguide into two narrow outputs in a 2.4 x 10 um2 footprint, with measured crosstalk down to -25 dB.","lead":"This paper builds and measures a tiny silicon photonic chip component that splits two wavelengths of light from a wide input channel into two separate narrow output channels without a long tapering section. The result is an ultra-compact wavelength demultiplexer that could shrink photonic circuits and connect directly to standard on-chip grating couplers.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Grating-coupler calibration is not described, so the reported -1.77/-2.10 dB insertion losses and crosstalk values cannot be separated from coupler and source response; the measured device performance is therefore not yet verified.","rationale":"The strongest_claim is an experimental demonstration with quantitative insertion loss and crosstalk. The load-bearing assumption is that the measured spectra reflect the demultiplexer alone. The manuscript's Characterization section omits any reference normalization, while the use of a supercontinuum source and grating couplers makes such normalization necessary. This is the same weakness identified by the reader, and it is concrete and addressable: a straight-waveguide reference on the same chip would settle it. The simulation-vs-experiment peak shift and the single-device measurement are secondary; neither changes the conclusion that the paper should be CONDITIONAL pending the calibration evidence. I find no independent reason to move the verdict.","tokens_in":9603,"tokens_out":7046,"duration_ms":79066,"concrete_test":"Require the authors to provide the raw measured spectra together with a reference-normalized measurement: a straight waveguide (or cut-back set) with the same input/output photonic-crystal grating couplers on the same chip, measured with the same source and analyzer. Subtract the reference fiber-to-fiber spectrum from the demultiplexer channel spectra and recompute peak transmission and crosstalk. If the calibrated values differ from the reported -1.77/-2.10 dB by more than about 1 dB, or crosstalk by more than about 3 dB, the headline experimental claims must be revised.","verdict_should_be":"UNCHANGED","load_bearing_attack":"To support the claim of an experimentally demonstrated low-loss demultiplexer, the measured fiber-to-fiber spectra must be attributable to the device rather than to the measurement apparatus. The Characterization section states only that 'a series of fully etched photonic crystal grating couplers were fabricated at each input and output port for chip-fiber coupling'; it does not describe a reference waveguide, cut-back set, or normalization to the supercontinuum source spectrum. Both the source and the grating couplers are strongly wavelength-dependent, so raw spectra cannot be interpreted as device S-parameters. This matters because the headline evidence is the set of dB values: -1.77/-2.10 dB insertion loss and -25.17/-12.14 dB crosstalk. The lower-channel crosstalk is already 11 dB worse than simulation (-23.50 dB), so even the direction of the discrepancy could be affected by coupler response. The single-device measurement with no error bars compounds the issue but is secondary. The concern is not that the authors misreported deliberately; rather, the manuscript as written does not establish that the quoted numbers are device-only values.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript reports the inverse design, fabrication, and characterization of a two-channel focusing wavelength demultiplexer based on regularized digital metamaterials. The design pipeline first uses the objective-first method to obtain a continuous permittivity distribution and then applies a modified steepest descent method with projection and low-pass filtering to enforce binarization and a minimum feature size of 120 nm. The resulting 2.4 μm × 10 μm device is designed to split 1520 nm and 1580 nm light from a 10-μm-wide input waveguide into two 0.48-μm-wide output waveguides, eliminating the need for an adiabatic taper. The authors report measured insertion losses of -1.77 dB and -2.10 dB and crosstalk of -25.17 dB and -12.14 dB for the upper and lower channels, respectively, and a numerical tolerance analysis claiming robustness to ±20 nm etch-depth and ±10 nm plane-size variations. A three-channel demultiplexer is also designed and simulated but not fabricated.","tokens_in":9814,"tokens_out":5284,"duration_ms":57540,"significance":"If the experimental results are taken at face value, the paper demonstrates that regularized inverse-designed digital metamaterials can realize an ultracompact wavelength demultiplexer that directly interfaces with a grating coupler without a taper, which would be a useful contribution to high-density silicon photonics. The optimization pipeline, including the minimum feature size constraint and the combination of objective-first initialization with steepest-descent refinement, is technically interesting and the three-channel numerical extension indicates potential scalability. However, the experimental validation as presented is incomplete: the measured spectra are not normalized to any reference, so the quoted insertion loss and crosstalk values cannot currently be separated from the grating-coupler and source response. The numerical tolerance study is also based on a restrictive uniform-perturbation assumption.","major_comments":[{"comment":"The measured spectra in Figure 4(b) appear to be raw fiber-to-fiber transmission through the input grating coupler, the device, and the output grating couplers. The Characterization section states only that photonic crystal grating couplers were fabricated at each port for chip-fiber coupling; it does not describe any reference waveguide, cut-back set, or normalization to the supercontinuum source spectrum. Because both the source and the grating couplers are strongly wavelength-dependent, the quoted insertion losses (-1.77 dB and -2.10 dB) and crosstalk values (-25.17 dB and -12.14 dB) cannot be attributed to the demultiplexer alone. The authors should provide reference-normalized measurements (e.g., a straight waveguide with identical grating couplers) or clearly restate the reported values as fiber-to-fiber quantities. Without this, the central experimental claim is not supported.","section":"Methods, Characterization; Figure 4"},{"comment":"The measured lower-channel crosstalk of -12.14 dB is more than 11 dB worse than the simulated value of -23.50 dB, whereas the upper channel changes by only about 1.5 dB. The paper attributes this discrepancy to RIE lag under-etching and to the fact that the 1565 nm field is spread throughout the design region, but this explanation is qualitative and no simulation of the estimated under-etched geometry is shown to reproduce the measured spectrum. To make the fabrication-imperfection argument load-bearing, the authors should include a simulated S-parameter response for an under-etched device (e.g., the 200 nm etch-depth case from Figure 5(a)) and directly compare the resulting lower-channel crosstalk with the measurement.","section":"Experimental Results and Analyses; Figure 4"},{"comment":"The design targets in the optimization specify output powers in the range [0.9,1] and [0,0.01] for the two channels, but the final regularized device with R=3 (Figure 4(a)) exhibits simulated peak transmissions of -1.39 dB (≈72.6%) and -1.45 dB (≈71.6%), which are outside the stated [0.9,1] interval. The manuscript does not explain this discrepancy or quantify how much of the performance loss is introduced by the binarization and minimum-feature-size constraints. The authors should report the performance of the unregularized continuous design alongside the regularized design and clarify whether the constraints are soft targets.","section":"Optimization Model Construction; Figure 4(a)"},{"comment":"The robustness analysis assumes that the etch depth and the plane size of all air holes vary uniformly. In practical RIE lag, the etch depth depends on the local feature size and pattern density, so a uniform shift is not necessarily the worst case and may overstate the claimed ±20 nm/±10 nm tolerances. The paper should justify the uniform-variation model or simulate a representative non-uniform perturbation, such as smaller holes being under-etched more than larger ones, to support the conclusion that the device is robust to fabrication errors.","section":"Tolerance to Fabrication Errors; Figure 5"}],"minor_comments":[{"comment":"The expression for the mode conversion efficiency appears garbled: the term '†c' is not defined and the integral notation is incomplete. Please rewrite this equation with clear definitions of all symbols.","section":"Equation (2)"},{"comment":"In the final sentence, 'Benefit From' should be 'Benefiting from' for grammatical correctness.","section":"Abstract"},{"comment":"The sentence 'The image show that all of the designed air holes were precisely reproduced' contains a subject-verb agreement error; it should be 'The image shows'. More importantly, a quantitative overlay of the SEM image with the designed layout would substantiate this claim of precise reproduction.","section":"Experimental Results and Analyses"},{"comment":"Reference [41] is cited as 'Under Review'; if this work is unpublished, the authors should provide a preprint identifier or remove the reference and rely on the discussion in the text.","section":"References"},{"comment":"The measured spectra in Figure 4(b) appear to cover a limited wavelength range; the authors should specify the scan range, resolution, and the number of devices measured, and include error bars or at least state the measurement repeatability.","section":"Figure 4"}],"recommendation":"major_revision","confidential_remarks":"The main concern is the missing calibration of the measured spectra. This appears to be an omission in the manuscript's description rather than a deliberate misreport, but it is load-bearing for the experimental claim. If the authors can supply reference-normalized data or a clear normalization procedure, the paper would be substantially strengthened. The design methodology and numerical studies are otherwise of interest for the photonic-device community, but the experimental demonstration as currently documented does not support the quoted insertion-loss and crosstalk figures."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague, here's the short version: this paper demonstrates an inverse-designed two-channel focusing wavelength demultiplexer on a 2.4 x 10 um2 footprint that connects directly to a 10-um-wide input without a taper. That geometry is new, and the simulated and measured results are in reasonable agreement. The experimental section, however, never describes how the grating-coupler response was calibrated out, so the quoted device-level -1.77/-2.10 dB loss and -25.17/-12.14 dB crosstalk are not yet established.\n\nWhat's actually good: the focusing concept is a real step beyond the narrow-input demux of Piggott 2015 and Su 2017. Removing the adiabatic taper by designing the device to accept a 10-um-wide mode is practically useful and non-obvious. The design flow—objective-first for a good starting point, then steepest descent with projection and a low-pass filter for binarization and minimum feature size—is well explained and produces clean, fabricable layouts. The measured redshift and the lower-channel degradation are plausibly explained by RIE lag, and the tolerance simulation gives a sensible if optimistic picture. The three-channel numerical example demonstrates extensibility.\n\nWhere it's soft: the calibration issue is the big one. The Characterization section mentions only that photonic crystal grating couplers were fabricated at each port; no reference waveguide, cut-back, or source normalization is described. The supercontinuum spectrum and the couplers are both strongly wavelength-dependent, so the raw spectra cannot simply be read as device S-parameters. This isn't an accusation—the authors may well have normalized and just didn't write it down—but as written the paper doesn't support the precision of the dB claims. The lower-channel crosstalk is already 11 dB off simulation, so even the discrepancy could be influenced by coupler response. Minor issues: measurements come from a single device with no error bars, and no layout or code is provided for exact reproduction. The tolerance analysis uses uniform perturbations, so it probably overstates robustness to non-uniform fabrication error.\n\nWho should read it: inverse-design photonics people and anyone working on compact silicon WDMs. The idea deserves discussion and the method is worth seeing. My recommendation: send it to peer review. The calibration question is fixable with a paragraph and maybe a plot, and the focusing geometry is a legitimate advance. If the authors can show the normalization, I'd be happy to see it published.","headline":"A genuinely new focusing WDM geometry with promising measured performance, but the missing grating-coupler calibration means the headline insertion-loss and crosstalk numbers aren't yet trustworthy.","tokens_in":10429,"tokens_out":2628,"would_cite":true,"duration_ms":27887,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A fabricated two-channel wavelength demultiplexer with a 2.4×10 μm² footprint separates 1520 nm and 1580 nm light with measured peak insertion loss of −1.77 dB and −2.10 dB.","keywords":["silicon photonics","wavelength demultiplexer","digital metamaterials","inverse design","objective first method","fabrication tolerance","focused devices","on-chip wavelength division multiplexing"],"falsifier":"Re-measure the fabricated chip against an on-chip reference: a straight 10-micron-wide waveguide with the same grating couplers at both ends. If the difference between demultiplexer transmission and reference transmission does not reproduce the reported −1.77 dB and −2.10 dB peaks with the stated crosstalk, then the missing grating-coupler calibration is the cause and the device-level numbers need revision.","tokens_in":9360,"feed_emoji":"🔀","tokens_out":8690,"duration_ms":88196,"temperature":0.7,"pith_summary":"This paper reports a two-channel on-chip wavelength demultiplexer, built from regularized digital metamaterials, that fits in 2.4 by 10 square microns and needs no long adiabatic taper. The design takes 1520 nm and 1580 nm light from a 10-micron-wide input waveguide and routes each wavelength into its own 0.48-micron-wide output waveguide. The fabricated device shows measured peak transmissions of −1.77 dB and −2.10 dB with crosstalk of −25.17 dB and −12.14 dB. The paper also simulates a three-channel version, arguing that the design method extends to other compact 'focused' devices. A sympathetic reader would care because this points to dense photonic circuits where wavelength splitting occupies a few tens of square microns rather than hundreds of microns of taper plus device area.","feed_headline":"A 24-µm² silicon chip separates two light wavelengths with ~2 dB loss","feed_subtitle":"Regularized inverse design routes 1520 nm and 1580 nm light into separate waveguides without long tapers.","key_machinery":"The load-bearing mechanism is a regularized digital metamaterial: a design region discretized into silicon/air pixels whose permittivity distribution is optimized to perform mode conversion. The optimization is a two-stage pipeline: an objective-first stage with ADMM produces a continuous permittivity distribution that largely satisfies the design targets, and a modified steepest descent stage applies a projection scheme to binarize the pixels and a low-pass filter enforcing a minimum feature size (here R=3, equivalent to 120 nm). The device performance is defined by overlap integrals between input and output modes, so the optimizer directly maximizes the power routed from the 10-μm input mode to the fundamental TE modes of the two output waveguides at their target wavelengths.","core_discovery":"The central claim is that inverse-designed digital metamaterials can be made both fabricable and high-performing if the optimization enforces binarization and a minimum feature size, and that this enables a 'focused' geometry in which a wide input waveguide is demultiplexed directly into narrow output waveguides. The demonstrated two-channel demultiplexer splits 1520 nm and 1580 nm light from a 10-μm input into two 0.48-μm outputs, with measured peak transmission of −1.77 dB in the upper channel and −2.10 dB in the lower channel, and crosstalk of −25.17 dB and −12.14 dB. The paper attributes the gap between simulated and measured spectra to etching imperfections, especially under-etching of small holes, and reports fabrication-tolerance simulations that allow ±20 nm in etch depth and ±10 nm in plane size. It also presents a simulated three-channel focused demultiplexer with crosstalk below −17 dB.","pith_inferences":["Editorial inference: the paper does not describe a reference calibration for the grating couplers used in the characterization, so the reported −1.77/−2.10 dB values should be read as device-plus-coupling numbers until such a normalization is shown; the true on-chip loss could be lower or higher depending on the couplers' behavior.","Editorial inference: the three-channel device is promising, but its real test is fabrication; a fabricated three-channel version with the same calibration would directly show whether the objective-first initialization remains effective as the number of output ports grows.","Editorial inference: the etching-depth tolerance analysis assumes all holes vary uniformly; a more demanding test would vary hole sizes independently, which is closer to what real lithography and etching produce."],"forward_implications":["If the measured performance is accurate, a wavelength demultiplexer can occupy about 24 square microns and attach directly to a 10-micron grating coupler, eliminating the hundreds-of-microns adiabatic taper normally needed.","The same regularized optimization pipeline should be reusable for other compact 'focused' components, such as mode splitters, mode converters, and power splitters.","Enforcing a 120-nm minimum feature size makes the inverse-designed pattern compatible with standard fabrication, which matters for moving such devices from simulation to production.","The simulated three-channel device suggests the approach is not limited to two wavelengths, although no three-channel fabrication result is reported."],"supporting_citations":[{"why":"Establishes the objective-first computational design framework that produces the continuous initial permittivity distribution.","marker":"[4]"},{"why":"Provides the inverse-designed broadband wavelength demultiplexer baseline and the mode-conversion-efficiency formulation used to define design targets.","marker":"[5]"},{"why":"Supplies the adjoint-gradient method used in the discrete steepest-descent optimization.","marker":"[11]"},{"why":"Prior digital-metamaterial ultracompact waveguide taper that this work replaces by a focused geometry without a taper.","marker":"[24]"},{"why":"Documents why 10-micron-wide grating couplers normally require long adiabatic tapers, the size problem this device avoids.","marker":"[29]"},{"why":"Supplies the projection and minimum-length-scale techniques used to binarize and regularize the permittivity distribution.","marker":"[32]"},{"why":"Gives the broader fabrication-constrained inverse-design approach that motivates the regularized digital metamaterial formulation.","marker":"[36]"}],"fun_headline_variants":["Inverse design yields 24-µm² two-wavelength demultiplexer","Chip-scale demux: 24 µm², −1.8 dB loss, no taper","Two-wavelength splitter fits on 24 µm² with 2 dB loss","No-taper optical demux: 24 µm², −1.8 dB insertion loss","Regularized metamaterials pack two-channel demux into 24 µm²"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported insertion-loss and crosstalk values are only trustworthy as device-level figures if the transmission of the input and output grating couplers was measured and divided out, and the characterization section does not describe such a normalization.","fun_headline_variants_meta":{"raw":{"variants":["Inverse design yields 24-µm² two-wavelength demultiplexer","Chip-scale demux: 24 µm², −1.8 dB loss, no taper","Two-wavelength splitter fits on 24 µm² with 2 dB loss","No-taper optical demux: 24 µm², −1.8 dB insertion loss","Regularized metamaterials pack two-channel demux into 24 µm²"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000651,"raw_usage":{"total_tokens":3036,"prompt_tokens":1046,"completion_tokens":1990,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":662,"completion_tokens_details":{"reasoning_tokens":1878}},"tokens_in":662,"tokens_out":1990,"duration_ms":15034,"temperature":1.0,"reasoning_tokens":1878,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T14:04:24.659679+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-measure the fabricated chip against an on-chip reference: a straight 10-micron-wide waveguide with the same grating couplers at both ends. If the difference between demultiplexer transmission and reference transmission does not reproduce the reported −1.77 dB and −2.10 dB peaks with the stated crosstalk, then the missing grating-coupler calibration is the cause and the device-level numbers need revision.","supporting_citations":[{"cited_title":"Nanophotonic Computational Design","cited_arxiv_id":null,"evidence_quote":"Establishes the objective-first computational design framework that produces the continuous initial permittivity distribution."},{"cited_title":"Y.; Lu, J.; Lagoudakis, K","cited_arxiv_id":null,"evidence_quote":"Provides the inverse-designed broadband wavelength demultiplexer baseline and the mode-conversion-efficiency formulation used to define design targets."},{"cited_title":"W.; Minkov, M.; Williamson, I","cited_arxiv_id":null,"evidence_quote":"Supplies the adjoint-gradient method used in the discrete steepest-descent optimization."},{"cited_title":"Adiabatic and Ultracompact Waveguide Tapers Based on Digital Metamaterials","cited_arxiv_id":null,"evidence_quote":"Prior digital-metamaterial ultracompact waveguide taper that this work replaces by a focused geometry without a taper."},{"cited_title":"V.; Claes, T.; Schrauwen, J.; Scheerlinck, S.; Bogaerts, W.; Taillaert, D.; O'Faolain, L.; Thourhout, D","cited_arxiv_id":null,"evidence_quote":"Documents why 10-micron-wide grating couplers normally require long adiabatic tapers, the size problem this device avoids."},{"cited_title":"S.; Wang, F.; Sigmund, O","cited_arxiv_id":null,"evidence_quote":"Supplies the projection and minimum-length-scale techniques used to binarize and regularize the permittivity distribution."},{"cited_title":"Y.; Petykiewicz, J.; Su, L.; Vučković, J","cited_arxiv_id":null,"evidence_quote":"Gives the broader fabrication-constrained inverse-design approach that motivates the regularized digital metamaterial formulation."}],"review_version":1}