{"id":"833fe8f5-bcf6-4d0e-8065-6745c9b6ba08","arxiv_id":"2504.12133","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"EUV scatterometry measured a copper interconnect dishing depth of 1.60 +/- 0.05 nm, and Fisher information analysis predicts optimal sensitivity near 14 nm wavelength at specific incidence angles.","lead":"This paper measures the depth of tiny recessed copper pads in a periodic semiconductor test structure using extreme ultraviolet scatterometry, reporting a dishing depth of 1.60 +/- 0.05 nm. It also uses Fisher information calculations to predict that a wavelength near 14 nm and carefully chosen incidence angles would make such measurements most sensitive.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Fitted dishing depth may be biased by the fixed carbon-contamination layer; the reported ±0.05 nm excludes systematic model error.","rationale":"The paper's strongest experimental claim is the 1.60±0.05 nm dishing depth measurement. I examined the FIM-based optimization and the noise model; the normalization-dependence on prior ranges is explicitly acknowledged in Section 5 and is a limitation of the secondary claim, not an internal flaw. The more serious issue is the forward-model uncertainty in the primary measurement. Section 3 introduces a fixed carbon layer to model contamination but provides no justification of its thickness, no uncertainty, and no sensitivity analysis. The reported ±0.05 nm comes from Monte Carlo resampling of the measured diffraction efficiencies and therefore covers only random CCD noise. Since EUV at 29 nm is strongly absorbed by carbon, the contamination layer and the shallow dishing feature both modify the top-surface phase and amplitude and can be strongly correlated. A systematic error in the carbon layer could shift the fitted dishing by more than the reported statistical uncertainty. This is testable by re-fitting with the carbon thickness varied. The paper otherwise has strengths: the noise model is analytic, the FIM is a standard and appropriate tool, and the Monte Carlo uncertainty is clearly described. My concern does not invalidate the methodology, but it makes the central measurement claim conditional on an unvalidated contamination model, which is exactly the condition that should be verified before accepting the numerical value as reported.","tokens_in":13724,"tokens_out":12720,"duration_ms":114238,"concrete_test":"Re-run the RCWA reconstruction with the carbon-layer thickness set to 0, 0.5, 1.0, 1.5, and 2.0 nm (and, if practical, with a ±10% change in its optical constants), holding all other model choices fixed, and record the best-fit dishing depth in each case. If the fitted dishing depth moves by more than the stated ±0.05 nm, the reported uncertainty is not a valid error budget and the central metrology claim requires an independent validation, such as AFM or TEM.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central experimental claim—a dishing depth of 1.60±0.05 nm—rests on an RCWA forward model that is not validated at the level claimed. Section 3 states that a thin carbon layer of fixed thickness was added to account for debris from dicing, and Fig. 2(a) notes the experimental error bars are higher than expected because of this debris. The paper does not report the carbon layer thickness, its optical constants, or any sensitivity study of the fitted dishing depth with respect to that layer. The reported ±0.05 nm comes only from a Monte Carlo over CCD noise (Section 3), so it captures statistical precision, not model-scale systematic error. Because the measured dishing depth is only ~1.6 nm, a contamination layer of comparable or larger thickness could easily shift the fitted value by more than 0.05 nm through correlated changes in the diffraction-efficiency curve. This is a load-bearing gap: if the dishing estimate is not robust to the contamination model, the paper's claim of single-nanometer sensitivity and profilometry for this sample is not established. A direct check—varying the carbon layer thickness across a plausible range and re-fitting—would settle whether the 1.60 nm value is an artifact.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript reports an EUV scatterometry experiment on a 2D periodic Cu/SiCN interconnect structure using 29 nm high-harmonic-generation light. The authors fit an RCWA forward model to seven measured diffraction-efficiency values and extract an average copper dishing depth of 1.60 ± 0.05 nm, where the quoted uncertainty comes from a Monte Carlo propagation of the CCD-noise statistics. They then construct a normalized Fisher information matrix using a Gaussian likelihood approximation and a two-parameter model (dishing depth and global angle offset) to predict optimal measurement conditions, finding a wavelength of 14.1 nm and incidence angles of 15.3° and 17.3°. The paper claims single-nanometer sensitivity to out-of-plane features and presents the FIM-based optimization as a generally applicable tool for EUV scatterometry experimental design.","tokens_in":14075,"tokens_out":12960,"duration_ms":117923,"significance":"If the dishing-depth measurement and the predicted optimal conditions are both quantitatively reliable, the work demonstrates a fast, non-destructive metrology relevant to semiconductor interconnect process control and provides a principled framework for selecting experimental parameters in coherent EUV scatterometry. The paper is commendable for explicitly deriving the Gaussian-likelihood FIM, including a camera-noise model, and for testing the approach on an industrially relevant sample. The experimental sensitivity curves in Fig. 2(a) usefully illustrate the single-nanometer response of the diffraction efficiency to the dishing depth. The supplemental derivation of the FIM from the log-likelihood is clear and standard.","major_comments":[{"comment":"The RCWA model includes a 'thin carbon layer of fixed thickness' to account for debris from dicing, but the thickness is not reported and no sensitivity analysis of the fitted dishing depth to this layer is given. Because the reported dishing depth is only about 1.6 nm, a contamination layer of comparable or larger thickness could shift the fitted value by more than the quoted ±0.05 nm, which reflects only Monte Carlo sampling of CCD noise, not model-scale systematic error. The authors should state the carbon layer thickness and its optical constants, and they should either vary the thickness over a plausible range and re-fit or treat it as an additional free parameter to quantify the resulting systematic uncertainty.","section":"Sec. 3"},{"comment":"Equation (1) and its Supplemental derivation appear to contain an incorrect error-propagation formula for the diffraction efficiency. For DE = 2N_±/(2N_± + N_0) with independent Poisson counts, the delta method gives σ²_DE = DE² [ (N_0²/N_±²) σ²_N±/(2N_± + N_0)² + σ²_N0/(2N_± + N_0)² ], but Eq. (1) instead contains σ²_N±/N_±² as the first term and an extra 4σ²_N± in the second term. Since this variance is used to compute the FIM elements in Section 4, the predicted optimal wavelength (14.1 nm) and incidence angles (15.3°, 17.3°) may be quantitatively incorrect. The authors should re-derive the noise model and recompute the optimization, or they should justify the approximation they used.","section":"Eq. (1), Sec. 4, and Supplemental"}],"minor_comments":[{"comment":"The abstract states that the experiment used λ = 29 nm light, while Section 2 says the multilayer mirrors are designed for a peak reflectance at 29.4 nm and Fig. 2(b) gives a nominal wavelength of 29.5 nm; please make these values consistent throughout.","section":"Abstract and Sec. 2"},{"comment":"The caption says the uncertainty is 'only .5 Å'; please use '0.5 Å' or, equivalently, '0.05 nm' to avoid ambiguity and to match the notation used in the main text.","section":"Fig. 3 caption"},{"comment":"The optimized experimental design is computed for a two-parameter model (dishing depth and angle offset), whereas the experimental reconstruction in Section 3 used four free parameters (also including SiCN density and wavelength offset); the authors should state explicitly whether the optimal conditions remain near-optimal when the additional parameters are included, or note this as a limitation.","section":"Sec. 4"},{"comment":"The global optimization of the FIM eigenvalue is not described in terms of the specific numerical algorithm used (e.g., genetic algorithm, multistart, or local optimization with random restarts); please provide this detail to clarify the robustness of the reported optimum.","section":"Sec. 4"}],"recommendation":"major_revision","confidential_remarks":"The paper fits the journal's scope and the experimental demonstration is interesting, but the two major issues above need to be addressed: the systematic error from the carbon contamination layer directly affects the credibility of the headline dishing-depth value, and the apparent error in the noise-model variance formula (Eq. 1) threatens the quantitative validity of the FIM optimization. The latter is particularly important because the optimization is a central advertised contribution, and even a corrected noise model may change the predicted optimal wavelength and angles."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is worth a careful read. It does two things: measures copper dishing depth on a 2D interconnect structure using 29 nm HHG scatterometry, and applies Fisher information matrix optimization to choose wavelengths and incidence angles. The experimental result is plausible, and the FIM part is a clean, well-documented application of a standard method.\n\nThe measurement itself is new and industrially relevant. The noise model is thoughtfully built, and the authors are honest about limitations: they acknowledge the optimized wavelength depends on prior ranges, and they explicitly flag the carbon contamination layer as a processing artifact. The FIM derivation is standard Gaussian likelihood math, and the optimization criterion (maximizing the minimum eigenvalue of the normalized FIM) is appropriate for their reconstruction approach.\n\nThe main soft spot is the uncertainty on the headline number. The reported ±0.05 nm comes only from Monte Carlo over CCD noise. The RCWA model includes a fixed carbon layer, but the paper gives neither its thickness nor its optical constants, and no sensitivity study with respect to that layer. Since the dishing depth is only ~1.6 nm, a contamination layer of comparable thickness could easily shift the fitted value by more than statistical error. This is a load-bearing gap for the claim of single-nanometer accuracy, though it is addressable: a simple study varying carbon thickness and re-fitting would settle it. A second, somewhat weaker limitation is that the FIM-optimal conditions (14.1 nm, 15.3° and 17.3°) are predictions only; they are not validated by a measurement at those settings. The data are also not public, which makes independent checking harder.\n\nWho is this for? People working in EUV metrology, scatterometry, or semiconductor process control. The FIM methodology is transferable, and the experimental dataset is a useful benchmark. The paper deserves a serious peer review; the carbon-layer robustness question should be raised with the authors, and the prior ranges in the FIM normalization deserve clearer reporting. I would not desk-reject it.","headline":"Solid experimental demonstration of EUV scatterometry for dishing metrology with a useful FIM design recipe; the reported uncertainty misses model-scale systematics.","tokens_in":14559,"tokens_out":1697,"would_cite":true,"duration_ms":19974,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":null,"created_at":"2026-08-16T12:36:59.415731+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":null,"supporting_citations":[],"review_version":1}