{"id":"08740d39-1ead-4f5c-a2f9-ff661e93dd07","arxiv_id":"2506.03372","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A genetic algorithm paired with an ABCD transmission-line model automatically synthesizes CPS-integrated planar THz band-stop filters at 0.6, 0.8, and 1.0 THz, with FEM validation.","lead":"This paper uses a genetic algorithm with a fast circuit model to automatically design tiny on-chip terahertz filters, then checks the results with a high-accuracy simulation. A generalist should care because it shows how to design custom THz components without slow trial-and-error simulations, potentially speeding up terahertz chip development.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The ABCD surrogate's mapping from arbitrary binary pixels to a canonical CPS cross-section (Eq.","rationale":"The reader's weakest assumption already identifies the quasi-static CPS transmission-line model as load-bearing, and I agree with that. My concern sharpens one specific aspect: the paper never defines how an arbitrary binary column maps to the W and S used in Eq. 7. The GA representation is a 2D binary grid, and the only stated constraint is continuity of a conductive pathway for DC biasing. That constraint does not guarantee a canonical two-strip CPS cross-section. If a column contains more than two metal regions (or a single region that cannot be decomposed into two strips), the characteristic impedance formula is not applicable and the surrogate's prediction is meaningless. The FEM comparisons in Fig. 4(c) and Fig. 7 are encouraging for the specific designs shown, but they are qualitative and concentrated at lower frequencies; the paper itself notes divergence above 0.9 THz, yet a 1.0 THz filter is presented. This is an internally stated limitation that the authors do not resolve. I do not think this requires changing the reader's CONDITIONAL verdict: the paper is a reasonable proof-of-concept, but the condition should explicitly require (a) disclosure and verification of the W/S extraction rule, and (b) quantitative ABCD-versus-FEM agreement at each demonstrated center frequency, especially 1.0 THz. The supporting evidence that the surrogate is fast and that one design matches HFSS well is real, but it does not cover the breadth of geometries the optimizer can produce.","tokens_in":11267,"tokens_out":5096,"duration_ms":63389,"concrete_test":"Reproduce the three center-frequency designs from Fig. 7(b); for each column of the binary chromosome, extract W and S exactly as implemented in the code (request the mapping from the authors if not in the text) and verify that every column is a single-gap, two-strip cross-section. Then simulate the 1.0 THz design in HFSS and report the peak |S21| deviation and the RMSE between the ABCD and HFSS results over 0.5-1.2 THz. If any column has more than two metal regions, or if the ABCD-HFSS error at the 1.0 THz stopband exceeds a few dB, the surrogate is not a reliable forward model for the generated designs and the framework is not validated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim depends on the surrogate (Eqs. 5-8) being a faithful forward model for the geometries the GA can generate. Eq. 7 is the closed-form impedance of a canonical coplanar-strip cross-section with a single gap S and two strips of width W. But the design representation is a 200-column by 75-row binary grid; nothing in the paper constrains each column to have exactly two metal regions, nor specifies how W and S are extracted from an arbitrary binary column. A column could contain multiple disconnected metal islands, a single wide conductor, or a slot pattern; for such columns the variables W and S in Eq. 7 are undefined, and the cascade of Eqs. 5-6 does not describe the electromagnetic response. The FEM validation is performed on the final 0.8 THz design (Fig. 4c) and qualitatively on the frequency-tuned filters (Fig. 7b); it does not establish that the surrogate is accurate across the structures explored during optimization. The admitted divergence above 0.9 THz is especially relevant because one demonstrated filter targets 1.0 THz (Fig. 7b); if the surrogate mispredicts the 1.0 THz design, the central claim that FEM validates it is not supported. The paper should either constrain the GA to a two-strip CPS topology or specify and verify the W/S extraction for every column, and provide quantitative ABCD-versus-HFSS error metrics at the target frequencies.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents a genetic algorithm (GA)-based inverse design method for planar THz band-stop filters integrated in coplanar striplines (CPS). The GA optimizes a 200×75 binary pixel pattern, with each column evaluated as a quasi-static CPS transmission-line segment via the ABCD matrix method (Eqs. 5–8). The fitness is a weighted RMSE between simulated and target S21/S11 magnitude and phase. Final validation is performed with ANSYS HFSS. Proof-of-concept designs include band-stop filters at 0.6, 0.8, and 1.0 THz with a fixed 2000-µm footprint and variable rejection depths. The paper claims a computational speed-up exceeding three orders of magnitude, reports convergence over 10 independent runs, and identifies itself as the first ABCD-based GA inverse design for CPS-integrated THz filters.","tokens_in":11546,"tokens_out":5537,"duration_ms":57146,"significance":"If the framework is validated, it offers a practical route to automated synthesis of fabricable planar THz filters, with a transparent speed/accuracy trade-off. The manuscript's strengths include a concrete runtime comparison (40 minutes for an optimization run versus more than 2 hours per HFSS simulation), an explicit connectivity constraint for future experimental biasing, repeated (10-run) convergence statistics, and an external FEM check for several designs. The core limitation is that the surrogate-to-full-wave agreement is only qualitative and is acknowledged to degrade above 0.9 THz, which is exactly where one demonstrated filter operates; this tempers the significance until quantitative validation is provided.","major_comments":[{"comment":"The ABCD surrogate treats each 10-µm column as a uniform CPS transmission line characterized by a single conductor width W and gap S (Eq. 7), but the design representation is a 200×75 binary grid with no stated rule for extracting W and S from an arbitrary column. In particular, the paper does not constrain each column to contain exactly two metal regions, nor does it define how W and S are computed for columns with disconnected metal islands, a single wide conductor, or slot-like patterns. Because Eq. (7) is undefined for such columns, the surrogate used throughout the GA may not describe the actual electromagnetic response of the generated geometries. Please specify the extraction rule and verify that the optimized designs (e.g., Fig. 4d) conform to it, or add a constraint that guarantees a two-strip CPS topology per column.","section":"Section II-A and Section III, Eqs. (5)–(8)"},{"comment":"The text and caption state that the 0.8 THz design 'meets the specified rejection depth and bandwidth with high accuracy,' but the target rejection is −50 dB while the computed response achieves approximately −46 dB. That is a 4-dB discrepancy at the central claim's primary metric. Please rephrase to 'approximately meets' or report the achieved value and deviation from the target quantitatively.","section":"Fig. 4(a) and Fig. 4 caption"},{"comment":"The paper acknowledges that ABCD results 'begin to diverge' from HFSS above 0.9 THz due to radiation and edge diffraction, yet one demonstrated filter is centered at 1.0 THz and the ABCD-HFSS comparison for it is described only qualitatively as 'strong agreement.' No quantitative ABCD-versus-FEM error metric (e.g., RMSE in S21 magnitude over the passband and stopband) is reported. Because the optimization objective is RMSE against a target, a quantitative surrogate-error assessment is required to support the claim that FEM validates the design, particularly at 1.0 THz where the surrogate is least trustworthy. Please include frequency-resolved error metrics for all three center-frequency filters.","section":"Section III, Figs. 4(c) and 7(b)"}],"minor_comments":[{"comment":"The text says 'The loss function L = -RMSE' but the right-hand side of Eq. (1) is a positive weighted sum of RMSE terms. Please reconcile the sign convention and explain how the GA maximizes fitness while the loss is negative RMSE.","section":"Eq. (1)"},{"comment":"The figure caption is labeled '(e)' even though it is a standalone figure; remove the stray '(e)'.","section":"Fig. 6"},{"comment":"The connectivity constraint is described as maintaining 'a continuous conductive pathway' for biasing, but a CPS line has two separate conductors; please clarify whether the constraint is applied to both strips and how the GA initialization enforces it.","section":"Section II-A"},{"comment":"The claim of being 'the first application of GA-based inverse design using the ABCD matrix method for CPS-integrated THz filters' is presented without a systematic literature comparison; please provide a more precise statement of what exactly is new relative to Refs. [18]–[23].","section":"Conclusion"},{"comment":"The manuscript reports a speed-up of 'exceeding three orders of magnitude' but the numbers given (40 minutes for 24,000 ABCD evaluations versus more than 2 hours for one HFSS run) suggest a much larger speed-up; please state the effective per-evaluation cost to avoid under- or over-statement.","section":"Section III"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is within the scope of the journal and the inverse-design idea is promising. The main revision points are the missing W/S extraction rule, the overstatement of the -50 dB target, and the absence of quantitative surrogate-versus-FEM errors. These are addressable with additional analysis and should not require new experiments."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's the take: this is a solid engineering proof-of-concept, not a breakthrough. The genuinely new thing is combining GA inverse design with an ABCD-matrix surrogate for CPS-integrated planar THz filters, plus a connectivity constraint to keep DC bias paths intact. That combination is not in the cited metasurface/antenna inverse-design literature, and the framework does what it claims: it synthesizes band-stop filters at 0.6, 0.8, and 1.0 THz in a fixed footprint, with tunable rejection depth, and checks the final designs with HFSS. The reported speed-up of over three orders of magnitude versus full-wave simulation is plausible, and the ten-run convergence study is a nice touch that many applied papers skip.\n\nThe soft spots are real but fixable. The biggest one: the ABCD surrogate (Eq. 7) assumes a canonical two-strip CPS cross-section with a single width W and gap S, but the GA operates on a 200x75 binary pixel grid. The paper never says how an arbitrary column with multiple metal islands, branches, or slots is converted into a single W and S. The connectivity constraint guarantees a DC path, not that each column is a clean two-strip line. So the surrogate's fidelity across the structures the GA actually explores is unverified. This is not a manufactured flaw; the text is silent on the mapping.\n\nSecond, the paper admits ABCD diverges from HFSS above 0.9 THz, then demonstrates a 1.0 THz filter. The caption for Fig. 4 calls the design \"meets\" a -50 dB rejection when Fig. 4(a) shows about -46 dB. No quantitative ABCD-vs-FEM error metrics are reported. These inconsistencies are worth cleaning up before the method is called validated.\n\nThe stress-test note sharpens the first point correctly. I don't think the note overstates the issue. Without a defined mapping, a reader cannot reproduce the surrogate, and the central claim rests on that surrogate.\n\nWho should read this: applied THz device people, particularly those working on on-chip passive components. It deserves a serious referee—the idea is useful and the limitations are addressable—but I would not cite it until the mapping and quantitative validation are in. Recommend peer review with major revision.","headline":"A credible proof-of-concept for GA inverse design of CPS-integrated THz filters, but the paper needs to define the pixel-to-CPS mapping and provide quantitative validation before the framework is fully convincing.","tokens_in":12052,"tokens_out":2215,"would_cite":false,"duration_ms":25314,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A genetic algorithm, guided by a fast matrix surrogate, can automatically design planar terahertz band-stop filters from a target spectrum.","keywords":["terahertz filters","genetic algorithm","inverse design","coplanar stripline","ABCD matrix method","band-stop filter","system-on-chip","surrogate modeling"],"falsifier":"Fabricate one of the inverse-designed filters on the silicon-nitride membrane and measure its transmission and reflection. If the measured stopband depth, center frequency, or bandwidth deviates from the full-wave prediction beyond the stated tolerances, especially for the 1.0 THz filter where the surrogate is already acknowledged to diverge, then the ABCD-guided search is not reliably delivering its designed spectral response.","tokens_in":11052,"feed_emoji":"🧬","tokens_out":7904,"duration_ms":85488,"temperature":0.7,"pith_summary":"This paper tries to show that a genetic algorithm, paired with a fast analytical surrogate, can automatically design planar terahertz band-stop filters etched into a coplanar stripline. The surrogate replaces full-wave simulation with cascaded ABCD transmission-line matrices, cutting the cost of evaluating one candidate design by more than three orders of magnitude. The authors demonstrate filters centered at 0.6, 0.8, and 1.0 THz with 150 GHz target bandwidths, and filters with 10, 20, and 30 dB rejection depths, all within one fixed 2000-micrometer footprint. Full-wave simulation confirms the final designs, suggesting a practical path from target spectrum to fabrication-ready geometry.","feed_headline":"Genetic algorithm evolves THz filter layouts from target spectra","feed_subtitle":"ABCD surrogate is 1000 times faster; full-wave simulation confirms 0.6-1.0 THz band-stop filters in one footprint.","key_machinery":"The load-bearing object is the cascaded ABCD transmission-line model of the pixelated filter, in which each 10 µm column along propagation acts as a uniform quasi-static CPS segment. Per-column impedance is assigned from the local geometry via the closed-form elliptic-integral formula $Z_{\\mathrm{CPS}} = (120\\pi/\\sqrt{\\varepsilon_{re}})\\,K(k_{\\mathrm{CPS}})/K(k'_{\\mathrm{CPS}})$, $k_{\\mathrm{CPS}} = S/(S+2W)$, and the total two-port response is the matrix product of all column segments. This surrogate runs orders of magnitude faster than full-wave analysis, which is what turns a 200-individual, 120-generation genetic search into a roughly 40-minute optimization run instead of weeks of simulation.","core_discovery":"The central claim is that inverse design of CPS-integrated planar THz filters becomes computationally practical when a quasi-static ABCD-matrix model serves as the fitness evaluator inside a genetic search. Each of the 200 columns of the 300 µm by 2000 µm design grid is treated as a uniform CPS transmission-line segment, with characteristic impedance computed from the local conductor width $W$ and strip spacing $S$ through the closed-form elliptic-integral formula $Z_{\\mathrm{CPS}} = (120\\pi/\\sqrt{\\varepsilon_{re}})\\,K(k_{\\mathrm{CPS}})/K(k'_{\\mathrm{CPS}})$ with $k_{\\mathrm{CPS}} = S/(S+2W)$. Cascading the per-column matrices and converting to scattering parameters reproduces the target magnitude and linear phase closely enough to guide evolution, and the paper reports that the optimizer meets specified spectral targets, including roughly $-46$ dB rejection against a $-50$ dB target at 0.8 THz. Full-wave validation shows strong agreement with the surrogate below about 0.9 THz, with divergence above that attributed to radiation and edge diffraction. The paper further claims that the inverse-designed filters beat a conventional periodic filter baseline in length and rejection depth, and identifies this as the first GA-plus-ABCD design flow for CPS-integrated THz filters.","pith_inferences":["A further step beyond the present framework would be to replace the quasi-static surrogate with a corrected or hybrid model above 1 THz, where radiation and edge diffraction dominate; the paper's own data provide the benchmark for how much the surrogate needs correcting.","Because the phase error carried only 10 percent of the loss weight and the target phase was linear, the demonstrated designs are not a strong test of phase-sensitive optimization; a stringent dispersion-shaping objective would be a harder and more informative validation.","The pixel grid of 4 µm by 10 µm restricts the geometry family, so the reported non-intuitive layouts are optimal only within that coarse binary representation; a multi-resolution refinement could push rejection depths closer to target.","An experimental comparison between the inverse-designed filter and the conventional periodic baseline, both fabricated and measured on the same platform, would be the decisive test of the paper's claimed superiority."],"forward_implications":["Designers can specify a target spectrum and get a fabricable CPS-integrated geometry without manual parameter sweeps.","The 40-minute optimization time on a desktop CPU makes multi-target searches and repeated runs with different seeds practical.","The connectivity constraint preserves a DC bias path, so the synthesized filters are compatible with terahertz system-on-chip experiments that need to bias photoconductive switches.","Because the surrogate and full-wave results agree below about 0.9 THz, the framework is currently most trustworthy for sub-terahertz designs; higher-frequency designs need explicit full-wave re-verification.","The same optimization loop should extend to other planar devices such as couplers, reflectors, and absorbers if the ABCD model can represent their unit cells."],"supporting_citations":[{"why":"Supplies the genetic-algorithm methodology and the rationale for using evolutionary search on non-convex electromagnetic design landscapes.","marker":"[24]"},{"why":"Provides the rank-based tournament selection probability used in Eq. (2) for parent selection.","marker":"[25]"},{"why":"Experimental validation of the analytical coplanar-line model, supporting the claim that the surrogate is accurate enough to guide design.","marker":"[27]"},{"why":"Supplies the closed-form elliptic-integral formula for CPS characteristic impedance used as each column's local segment model.","marker":"[28]"},{"why":"Gives the standard ABCD-to-scattering-parameter conversion used to evaluate fitness from the cascaded matrices.","marker":"[29]"},{"why":"Provides the conventional periodic filter baseline and comparison numbers for the claimed length reduction and deeper rejection.","marker":"[11]"}],"fun_headline_variants":["GA inverse-designs THz filters with fast ABCD surrogate","Genetic algorithm crafts planar THz filters from target spectra","Inverse-designed THz band-stop filters from 0.6 to 1.0 THz","GA designs CPS THz filters in one footprint with tunable rejection"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that every 10-micrometer column of the binary pattern behaves as a uniform quasi-static CPS transmission-line segment whose impedance depends only on local width and spacing, with no radiation, edge diffraction, or coupling between columns; the paper itself reports that this model begins to diverge from full-wave simulation above about 0.9 THz, and one of the demonstrated filters is centered at 1.0 THz.","fun_headline_variants_meta":{"raw":{"variants":["GA inverse-designs THz filters with fast ABCD surrogate","Genetic algorithm crafts planar THz filters from target spectra","Inverse-designed THz band-stop filters from 0.6 to 1.0 THz","GA designs CPS THz filters in one footprint with tunable rejection"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000624,"raw_usage":{"total_tokens":2911,"prompt_tokens":987,"completion_tokens":1924,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":603,"completion_tokens_details":{"reasoning_tokens":1847}},"tokens_in":603,"tokens_out":1924,"duration_ms":17486,"temperature":1.0,"reasoning_tokens":1847,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T11:04:32.807913+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Fabricate one of the inverse-designed filters on the silicon-nitride membrane and measure its transmission and reflection. If the measured stopband depth, center frequency, or bandwidth deviates from the full-wave prediction beyond the stated tolerances, especially for the 1.0 THz filter where the surrogate is already acknowledged to diverge, then the ABCD-guided search is not reliably delivering its designed spectral response.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the genetic-algorithm methodology and the rationale for using evolutionary search on non-convex electromagnetic design landscapes."},{"cited_title":"Genetic algorithms, tournament selection, and the effects of noise,","cited_arxiv_id":null,"evidence_quote":"Provides the rank-based tournament selection probability used in Eq. (2) for parent selection."},{"cited_title":"Terahertz attenuation and dispersion characteristics of coplanar transmission lines,","cited_arxiv_id":null,"evidence_quote":"Experimental validation of the analytical coplanar-line model, supporting the claim that the surrogate is accurate enough to guide design."},{"cited_title":"Analytical formulas for coplanar lines in hybrid and monolithic MICs,","cited_arxiv_id":null,"evidence_quote":"Supplies the closed-form elliptic-integral formula for CPS characteristic impedance used as each column's local segment model."},{"cited_title":"Demonstration of a planar multimodal periodic filter at thz frequencies,","cited_arxiv_id":null,"evidence_quote":"Provides the conventional periodic filter baseline and comparison numbers for the claimed length reduction and deeper rejection."}],"review_version":1}