REVIEW 3 major objections 6 minor 10 references
Practical Evaluation of FFT-Based Thickness Extraction for Thick-Film Reflectometry
T0 review · 3 major / 6 minor · reviewed 2026-08-14 · deepseek-v4-flash
Pith's one-line read On a production 52 μm film, FFT-based thickness extraction is systematically off by 0.62 μm RMSE against a white-light interferometry reference, while zero-crossing fringe tracking reduces the error to 0.20 μm.
desk verdict Useful production-wafer comparison showing FFT can be biased despite good repeatability, but the WLI reference and the leakage explanation need stronger support. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The carrying mechanism is the two-beam reflectance spectrum $R(k_0) = R_0 + R_1 \cos(2 n_1 d_1 k_0 - \phi_0)$ written in wavenumber space, whose oscillation frequency is the optical thickness $2 n_1 d_1$. FFT extraction locates the dominant frequency of this oscillation, but the discrete Fourier transform assumes the measured window is one period of a periodic signal; when the instrument's fixed spectral window does not contain an integer number of fringes, the implied periodic continuation has discontinuities at the window edges, and spectral leakage shifts the peak. LRZ avoids this by locating zero crossings of the detrended spectrum and fitting the phase progression directly, so thickness is estimated from phase increments rather than from a localized Fourier bin.
What would settle it
On the same 52 μm film, acquire reflectance spectra with spectral windows that contain an exactly integer number of interference fringes (e.g., by adjusting the sweep endpoints) and recompute FFT thickness. If the 0.62 μm RMSE offset persists despite integer-period windows, spectral leakage is not the dominant error; if it disappears, the paper's mechanism is confirmed.
Extended reading notes
Core claim
The central claim is that FFT-based thickness extraction, although widely used for thick-film reflectometry and highly repeatable in practice, carries a systematic accuracy penalty that grows with fringe density. On a nominal 52 μm dielectric film, the FFT results deviate from the WLI reference with RMSE = 0.621 μm and correlation r = 0.884, while LRZ achieves RMSE = 0.204 μm (r = 0.985) and optical model fitting achieves RMSE = 0.093 μm (r = 0.997). The paper attributes the FFT error to spectral leakage caused by finite measurement windows and non-integer fringe periodicity: because the discrete Fourier transform implicitly assumes periodic continuation of the measured signal, truncating the reflectance spectrum at an arbitrary phase redistributes energy into neighboring frequency bins and shifts the apparent peak. Simulated signals with identical optical thickness but different numbers of fringes in the window confirm that non-integer truncation moves the FFT peak. As thickness increases, a fixed spectral window contains more fringes, making the peak localization more sensitive to this truncation effect, whereas LRZ benefits from the additional zero crossings for phase regression.
Load-bearing premise
The accuracy ranking assumes that WLI is a valid reference for a transparent 52 μm film; the paper relies on a cited prior study for that reliability and does not validate WLI against another absolute technique on this wafer.
Editorial extensions
If this is right
- If the paper is right, FFT-based thick-film thickness readouts in production settings should be treated as systematically biased, not just noisy; calibrating or abandoning them may be needed when absolute thickness matters.
- The bias should grow with film thickness on a fixed spectral window, so FFT's suitability degrades exactly in the thick-film regime where it is most often chosen for speed.
- LRZ offers a non-iterative, model-free middle path: it is only 0.11 μm worse than full model fitting on this test, while avoiding optical-constant knowledge and fitting cost.
- A production metrology workflow that reports only repeatability (e.g., standard deviation) can certify a method that is reproducibly wrong; agreement with an independent reference must be part of qualification.
Reading between the lines
- The spectral-leakage mechanism is not specific to FFT thickness extraction: any estimator that localizes a peak in a finite measurement window (e.g., spectral-domain OCT, white-light interferogram analysis) should exhibit a similar thickness-dependent bias whose magnitude scales with the number of fringes in the window.
- The paper's explanation yields a direct test: resampling or re-windowing the same spectra so that each window contains an integer number of fringes should collapse the FFT RMSE toward the LRZ value; if it does not, another error source (e.g., dispersion, phase offset) is at work.
- A correction curve could in principle be learned from the observed residual pattern and applied to FFT outputs, restoring speed while removing bias, but that would make FFT dependent on reference metrology and undercut its model-free appeal.
- The repeatability-versus-accuracy lesson generalizes to any inline metrology where the estimator has a systematic error that random noise does not reveal; reporting only gauge repeatability and reproducibility can mask such errors.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper compares three thickness-extraction methods—FFT, Linearized Reflectance Zero-Crossing (LRZ), and optical model fitting—applied to reflectometry spectra of a nominal 52 μm dielectric film on a production wafer, using white-light interferometry (WLI) as the reference. Nine sites were measured nine times each. FFT shows the best repeatability (σ < 0.04 μm) but the largest deviation from WLI (RMSE = 0.621 μm, r = 0.884); LRZ gives RMSE = 0.204 μm (r = 0.985) and model fitting RMSE = 0.093 μm (r = 0.997). The authors attribute the FFT bias to spectral leakage from finite measurement windows and non-integer fringe counts, supported by a simulation. The conclusion is that repeatability does not imply accuracy, and phase-tracking methods such as LRZ are a practical alternative.
Significance. If the reported comparison is correct, the paper provides a useful empirical demonstration that high repeatability does not guarantee accuracy for FFT-based thick-film reflectometry, and it offers an independent experimental comparison of three extraction methods on a production sample. The experimental design is reasonable: nine sites, nine repeats per site, and an external WLI reference. However, the strength of the accuracy claim depends on two points that are not fully established: the validity of WLI as a ground truth for this specific transparent film, and the quantitative link between the spectral-leakage simulation and the experimental conditions. With these gaps, the quantitative ranking should be treated as provisional.
major comments (3)
- [Experimental comparison (WLI reference)] The accuracy ranking is anchored entirely to WLI as ground truth, but the paper does not validate WLI on the actual sample. WLI thickness extraction for a transparent 52 μm film requires knowledge of the refractive index and correct handling of fringe-order ambiguity; a bias of only a few tenths of a micrometer would be small compared with the FFT RMSE but comparable to the LRZ RMSE, and could reorder the FFT/LRZ comparison. The sole support is a general citation (ref. 8). The authors should either report a WLI calibration or cross-check (e.g., step-height standard or comparison with model-fitting results) or provide an uncertainty budget for the WLI reference.
- [Fig. 3 / spectral leakage analysis] The spectral-leakage explanation is only qualitative. The simulation uses signals with 2.6, 3.0, and 3.4 fringe periods, but the actual fringe count for the 52 μm film over the NANOSPEC spectral window is not stated anywhere in the paper and is likely far larger than a few periods. The paper does not compute the expected FFT bias for the instrument's actual wavelength range and film thickness, nor compare that expectation to the observed 0.62 μm RMSE. Without this quantitative link, the statement that the deviation is 'consistent with spectral leakage' is not established. Please provide the instrument's spectral range and a simulation with the real parameters.
- [Fig. 2 / statistics] The RMSE values and correlation coefficients are computed from nine site averages, but no confidence intervals, standard errors, or significance tests are reported. With only nine points, the difference between FFT (RMSE 0.621 μm) and LRZ (RMSE 0.204 μm) may or may not be statistically significant. The repeatability data from the 81 measurements should be used to assess the uncertainty of the site averages, and the accuracy comparison should include appropriate error bars or a paired-test statistic.
minor comments (6)
- [Experimental setup] The manuscript does not specify the spectral range of the NANOSPEC 9100 or the WLI settings; please add this information so the fringe count and spectral-leakage calculation can be reproduced.
- [Model fitting] The optical model fitting procedure is not described: no model type, optical constants, or fitted parameters are reported. This omission hampers reproducibility of the best-performing method.
- [Discussion of leakage scaling] The statement 'As film thickness increases, more interference fringes are contained within a fixed spectral measurement window, making FFT peak localization increasingly sensitive to finite-window effects' is counterintuitive and needs a quantitative justification; higher fringe counts might be expected to improve frequency resolution.
- [Self-citation] The paper's references to refs. 6 and 7 are self-citations of the LRZ method; the authors should identify these as their own work and explain the LRZ algorithm briefly in the text for readers.
- [Fig. 3] Fig. 3 lacks axis labels and quantitative values; please state the simulated optical thickness, fringe counts, and the resulting thickness errors.
- [Typos] There are minor typographical issues: 'repeata bility' in the Fig. 1 caption, and inconsistent use of 'measurements' vs 'measurement' in the abstract.
Circularity Check
No circularity: empirical comparison against an external WLI reference; LRZ self-citations are background only.
full rationale
The paper's central empirical comparison is anchored to WLI, an external reference technique (Bruker Contour GT-X), supported by an external citation (ref. 8, G. Huang et al., Micromachines 2025), not by the authors' own prior work. No parameter is fitted to the WLI data and then reported as a prediction; the FFT, LRZ, and model-fitting thickness values are extracted from reflectometry spectra by their respective published algorithms and compared to WLI afterward. The spectral-leakage explanation is supported by standard external references (Quinten, ref. 3; Harris, ref. 10) and by the paper's own simulations with synthetic signals of known optical thickness, so the mechanism is not defined in terms of the measured result. LRZ is the authors' prior method (refs. 6-7), but the present paper does not rely on those papers for the quantitative RMSE and correlation values; those numbers come from the 81 measurements analyzed here. At most, the background description of LRZ's phase-tracking behavior cites previous work, but this is not load-bearing for the empirical ranking. The paper's reliance on WLI as ground truth for a transparent 52 micrometer film is asserted rather than validated in this manuscript; that is a correctness or validity risk, not a circularity, because WLI is an independent measurement principle and the cited support is external to the authors. No equation in the paper is equivalent by construction to the claimed conclusion, and no fitted input is renamed as a prediction.
Assumptions & free parameters
assumptions (4)
- domain assumption Two-beam interference model (Eq. 1) adequately represents the reflectance spectrum of the 52 μm film.
- domain assumption WLI provides an accurate and unbiased reference thickness for the film (ref 8).
- standard math Spectral leakage theory (Harris 1978) applies to the FFT processing used, including the specific window and zero-padding choices.
- domain assumption The LRZ algorithm implementation matches references 6 and 7 and performs zero-crossing detection as described.
Cite this review
Pith. "Pith review of Practical Evaluation of FFT-Based Thickness Extraction for Thick-Film Reflectometry." pith.science (2026). https://pith.science/paper/XJCPI7IB
@misc{pith2026260810146,
author = {Pith},
title = {Pith review of: Practical Evaluation of FFT-Based Thickness Extraction for Thick-Film Reflectometry},
year = {2026},
howpublished = {\url{https://pith.science/paper/XJCPI7IB}},
note = {Machine review of arXiv:2608.10146}
}
read the original abstract
Fast Fourier transform (FFT) is widely used for thick-film reflectometry because of its simplicity and computational efficiency. However, its performance under practical thick-film measurement conditions has received limited experimental evaluation. In this work, FFT, Linearized Reflectance Zero-Crossing (LRZ), and optical model fitting were compared using reflectometry measurements from a nominal 52 {\mu}m dielectric film acquired on a production wafer, with White-Light Interferometry (WLI) serving as an independent reference. Although FFT exhibited excellent repeatability, it showed systematic deviation from WLI (RMSE = 0.62 {\mu}m), whereas LRZ significantly improved accuracy (RMSE = 0.20 {\mu}m). Simulations and theoretical analysis indicate that the observed systematic deviation is consistent with spectral leakage caused by finite measurement windows and non-integer fringe periodicity. These results demonstrate that excellent repeatability does not necessarily imply high accuracy and highlight the advantages of zero-crossing-based approaches for thick-film thickness metrology.
Figures
Reference graph
Works this paper leans on
-
[1]
N. G. Orji, M. Badaroglu, B. M. Barnes, C. Beitia, B. D. Bunday, U. Celano, R. J. Kline, M. Neisser, Y . Obeng and A. E. Vladar, Nat. Electron. 1 [10], 532 (2018)
work page 2018
-
[2]
M. A. Taranov, B. G. Gorshkov, A. E. Alekseev, Yu. A. Konstantinov, A. T. Turov, F. L. Barkov, Z. Wang, Z. Zhao, M. S. D. Zan and E. V Kolesnichenko, Instruments and Experimental Techniques 66 [5], 713 (2023)
work page 2023
-
[3]
M. Quinten, SN Appl. Sci. [ DOI:10.1007/s42452-019-0866-9]
-
[4]
P. J. de Groot, Reports on Progress in Physics 82 [5], 056101 (2019)
work page 2019
-
[5]
Z. Zhou, E. A. Lopez-Guerra, B. Zhou, M. Kwan, P. Wilkens and C. Chien, Journal of Micro/Nanopatterning, Materials, and Metrology [ DOI:10.1117/1.JMM.24.2.024001]
-
[6]
Z. Zhou, E. A. Lopez-Guerra, I. Zana, V . Nguyen, N. Q. H. Tran, V . Huang, B. Zhou, G. Qian, M. Kwan, P. Wilkens and C. Chien, Metrology 6 [1], 13 (2026)
work page 2026
-
[7]
Z. Zhou, E. A. Lopez-Guerra, I. Zana, N. Q. H. Tran, B. Zhou, G. Qian, M. Kwan, P. Wilkens and C. Chien, in Metrology, Inspection, and Process Control XL, eds. H. Cramer and N. G. Schuch (SPIE, 2026) p. 82
work page 2026
- [8]
Show all 10 references
-
[9]
Angus Macleod, Thin-Film Optical Filters, Fifth Edition (CRC Press, 2017)
H. Angus Macleod, Thin-Film Optical Filters, Fifth Edition (CRC Press, 2017)
2017
-
[10]
F. J. Harris, Proceedings of the IEEE 66 [1], 51 (1978). Figures Fig. 1. Thickness measurements obtained from FFT, LRZ, and optical model fitting across nine wafer sites (top row), together with the corresponding standard deviation from nine repeated measurements at each site ...
1978
Reviewed August 14, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.