Spectrally smooth broadband response via autocorrelation-constrained inverse design

Johannes Gedeon; Ole Sigmund; Rasmus Elleb{\ae}k Christiansen

Spectrally smooth broadband response via autocorrelation-constrained inverse design

T0 review · 2 major / 6 minor · reviewed 2026-08-14 · deepseek-v4-flash

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

Pith's one-line read This paper claims that a weighted long-lag autocorrelation metric turns narrow in-band spectral defects, which integrated energy objectives miss, into a visible time-domain penalty, and that constraining it yields broadband dielectric…

desk verdict A useful, incremental tool paper: a time-domain autocorrelation metric that gives gradient-based broadband photonic design sensitivity to in-band ripple that energy-only objectives miss, deserving a normal peer-review round. read the letter →

arxiv 2608.13367 v2 pith:OX2POPYT submitted 2026-08-13 physics.optics

Johannes Gedeon , Rasmus Elleb{ae}k Christiansen , Ole Sigmund This is my paper · ORCID

classification physics.optics

keywords autocorrelationmetricbroadbandreflectanceinversedesigntopologyoptimizationFDTDtime-domainadjointmethodBragggratingspectralsmoothness

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

Broadband inverse design in optics usually optimizes time-integrated energies, which by Parseval's theorem measure only the total response over a band and cannot see how that response is distributed in frequency. The paper proposes a time-domain metric $T_{AC}$, the weighted long-lag autocorrelation energy of the reflected field relative to the incident pulse, that is Fourier-dual to spectral sharpness and therefore exposes narrow dips, ripple, and sidelobes. Adding $T_{AC}\le 1$ as a constraint, or a weak penalty, removes these defects in one-dimensional dielectric mirrors while leaving median reflectance essentially unchanged: $0.8096$ unconstrained versus $0.8093$ constrained over 100 runs. If this works as claimed, designers get a cheap way to enforce spectral flatness in time-domain inverse design without prescribing a target waveform or adapting the excitation pulse.

What carries the argument

The central object is $T_{AC}[s;s_0] = E_{AC}[s]/E_{AC}[s_0]$, with $E_{AC}[s] = 2\int_{\tau_{\min}}^{\infty}(\tau/\tau_{\min})^2 |c_s(\tau)|^2\,d\tau$. It measures weighted long-lag autocorrelation energy of the reflected signal relative to the incident pulse; the reference delay $\tau_{\min}$ is the first lag where the Hilbert envelope of the reference autocorrelation has decayed to 1% of its peak. By the Fourier-dual relation between $\partial_f S$ and $\tau C(\tau)$, this machinery converts sharp spectral features into long autocorrelation tails that the optimizer can see, quantify, and penalize either as a constraint or a weighted objective.

What would settle it

Run the unconstrained optimization of Sec. III-A and re-evaluate $T_{AC}$ with progressively longer simulation times so a narrow phase-defect dip is resolved at finer frequency resolution; if the dip deepens while $T_{AC}$ stays near the reference value of 1, then the long-lag autocorrelation metric is insensitive to that class of spectral features.

Watch

Extended reading notes

Core claim

The central discovery is that the long-lag tail of the autocorrelation of the reflected field, weighted by lag squared and compared with a broadband reference pulse, is a computable time-domain surrogate for narrow-band spectral nonuniformity. Because the energy spectral density $S(f)=|\hat{s}(f)|^2$ and the autocorrelation $C(\tau)$ form a Fourier pair, a localized spectral feature produces a slowly decaying oscillatory tail; the quadratic weight makes the metric sensitive to $\partial_f S$, hence to sharp features and ripple. Adding $T_{AC}\leq 1$ to the topology-optimization problem removes a phase-defect-like resonance in the reflectance spectrum, and across 100 random initializations the constrained ensemble keeps median reflectance at $0.8093$ versus $0.8096$ unconstrained while reducing $T_{AC}$ outliers. A weak penalty formulation with weighting $\alpha=10^{-2}$ even reaches higher reflectance than the unconstrained problem while driving $T_{AC}$ near 1, and strong penalization recovers apodized and chirped grating morphologies whose near-unity reflectance matches a parametric linear-chirp reconstruction.

Load-bearing premise

The claim stands on the assumption that the weighted long-lag autocorrelation tail of the reflected signal, measured against the incident pulse's coherence time, faithfully captures every narrow in-band spectral defect that matters, so that suppressing that tail is enough to guarantee a spectrally smooth response.

Editorial extensions

If this is right

Maximizing the Parseval-type energy fraction alone is insufficient: one unconstrained optimization produced a reflectance spectrum with a narrow phase-defect-like resonance and $T_{AC}=8.628$ even at a local optimum.
Enforcing $T_{AC}\le 1$ removes such defects without statistically reducing reflectance, with median $\eta_r$ equal to 0.8093 against 0.8096 for the unconstrained ensemble.
A weak penalty with $\alpha=10^{-2}$ in $J=(1-\alpha)\eta_r-\alpha T_{AC}$ finds designs with higher reflectance than no penalty while keeping $T_{AC}$ close to 1.
Strong penalization trades reflectance for spectral flatness and yields apodized and chirped grating morphologies with near-unity flat reflectance, with $\eta_r=0.978$ for the chirped design.
The same metric can be used as a post-evaluation diagnostic and can be carried into other broadband inverse-design problems, including near-field coupling, achromatic focusing, and broadband absorption.

Reading between the lines

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

Editorial inference: because $T_{AC}$ is defined on the reflected signal alone, the same construction should apply to transmission or absorption spectra, making it a general-purpose spectral-flatness regularizer for any transfer function once a reference pulse is chosen.
Editorial inference: the reference delay $\tau_{\min}$ is derived from the incident pulse, so the metric's sensitivity can be tuned per spectral region; a natural extension is a multi-scale version with several threshold lags that targets ripple at different bandwidths simultaneously.
Editorial inference: the automatic emergence of a chirped grating suggests the metric could serve as an unsupervised probe for discovering structured dispersive designs, and comparing its Pareto front with hand-built apodized and chirped templates would isolate what the autocorrelation constraint adds beyond known grating recipes.

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

You should know this one before deciding what to do with it: it is a solid, incremental methods paper, not a field-reshaping result. The new thing is T_AC, a normalized weighted long-lag autocorrelation energy of the reflected field, introduced as a constraint or penalty in FDTD-based topology optimization. The underlying Wiener–Khinchin idea is textbook, and the authors say so. What is theirs is the specific definition, the normalization to the incident pulse, and the demonstration that it works on 1D Bragg mirror design.

Credit where it is due. The math in Sec. II is clean, the 100-run ensemble comparison is a real step beyond the usual single-start demo, and the median reflectance is essentially unchanged while T_AC drops. They also ship code and data, which is what makes the paper reviewable. The chirp reconstruction in Sec. III-C is a nice bonus, though not load-bearing.

Now the soft spots. The finite-window worry is real. With T_sim around 1.58 ps (25,000 steps at 63.38 as), any spectral dip narrower than roughly 0.5–1 THz has an autocorrelation tail that extends beyond the recorded window. The computed E_AC then undercounts that tail, so the constraint can be satisfied not because the spectrum is smooth but because the offending feature lives outside the measurement. The paper says T_sim is long enough for fields to decay but gives no convergence check, like E_AC or T_AC versus T_sim. In the demonstrated mirror case the visible feature is wide enough that the benefit still shows clearly, so this is a limitation of the method's claimed range, not a fatal flaw. I would ask for a supplementary T_sim sweep.

Two smaller things. The adjoint derivative of T_AC is described in prose, not written out; the code fills the gap but an explicit expression would help. And the weak-penalty conclusion in Sec. III-B rests on single-start runs, unlike the constraint result, which has the ensemble. That makes the Pareto story anecdotal. Both are minor.

Who is this for: people doing time-domain inverse design in photonics, particularly topology optimization. It merits a serious referee, and in fact would benefit from one who knows signal processing and FDTD—to push on the truncation issue and on how generally the metric maps to spectral quality. Yes, send it out.

Referee Report

2 major / 6 minor

Summary. The manuscript introduces a time-domain metric T_AC, defined as the weighted long-lag autocorrelation energy of the reflected field normalized by that of the incident broadband pulse (Eq. (3)), and argues that this quantity is sensitive to narrow in-band spectral features that conventional Parseval-type integrated objectives miss. The metric is incorporated into FDTD-based density topology optimization of 1D dielectric Bragg mirrors, either as a hard constraint (P_c, Eq. (6b)) or as a penalty term (Eq. (7)). Numerical results over 100 random initial designs show that the constrained ensemble exhibits visibly reduced sharp spectral features with essentially unchanged median reflectance (0.8096 vs. 0.8093), while the penalty formulation produces a Pareto trade-off and recovers apodized and chirped grating profiles (Sec. III). The paper includes FDTD and adjoint-method details in Appendix A and states that the code and dataset are publicly available.

Significance. If the central claim holds, the paper makes a useful contribution to time-domain inverse design, where quadratic integrated objectives are known to be blind to spectral ripple and localized defects. The Fourier derivation in Sec. II is clear and essentially correct, and the 100-run ensemble comparison is a genuine strength because it addresses the non-convexity of the design problem. The open code and dataset are also valuable for reproducibility. However, the finite FDTD recording window may compromise the very long-lag tails that T_AC is designed to measure, and the paper does not yet provide the convergence check needed to separate the effect of the constraint from a finite-window artifact. For this reason the demonstrated benefit is not yet fully established.

major comments (2)

[Appendix A, Table II and Eq. (3)] The finite FDTD recording window is not separated from the claimed effect of the constraint. For the runs in Figs. 4 and 5, T_Sim = 25,000 × 63.38 as ≈ 1.58 ps, so the FFT-based autocorrelation in Eq. (3) is evaluated only for lags |τ| < T_Sim and all longer lags contribute zero. A narrow spectral dip of width Δ ≈ 0.1 THz produces an autocorrelation tail that decays on a timescale of order 1/Δ ≈ 3.2 ps, longer than the window; such a defect would be essentially invisible to T_AC even though it is exactly the class of feature the metric is meant to detect. The same truncation affects η_r in Eq. (4), because late-time reflected energy is omitted. The statement in Appendix A that T_Sim is "sufficiently long for the fields to decay" is not supported by any reported convergence check. Please add a T_Sim-convergence study, for example by doubling T_Sim for representative constrained and unconstrained designs and reporting η_r, T_AC, and residual field energy, and verify that the sharp-feature suppression in Fig. 4(b.1) persists when the spectra are computed from longer runs.
[Section II and Appendix A] The metric's parameter choices, specifically the 1% Hilbert-envelope threshold defining τ_min and the quadratic lag weight in E_AC, are introduced without any sensitivity or robustness analysis. Since the central claim is that T_AC is a faithful proxy for narrow in-band spectral defects, the paper should demonstrate that the reported smoothing and the T_AC = 1 bound are not artifacts of these particular values. A minimal check would be to repeat one constrained optimization with τ_min varied by a factor of two and with a linear instead of quadratic lag weight, and to report how the optimized spectra and T_AC values change.

minor comments (6)

[Sec. III-A, Fig. 4(b.2)] The text reports the unconstrained median T_AC = 1.854 but does not quote the median or spread of T_AC for the constrained ensemble; please report these values so the reader can confirm that the constraint is actually satisfied across the ensemble.
[Sec. III-B] The expression "Mnd − → 0" appears to be a typographical error; it should read "M_nd → 0".
[Eq. (3) and surrounding text] The notation is inconsistent between T_AC and TAC; please use a single subscripted form consistently throughout the manuscript.
[Sec. II] Please state explicitly that E_AC is a heuristic estimator rather than an exact spectral-derivative norm, because the integral is restricted to |τ| > τ_min and uses a normalized autocorrelation; this would clarify the relationship between Eq. (3) and the Parseval-type identity that motivates it.
[Sec. III-C and Table I] The fitted chirp parameters a1, a2, and D_0 are reported without uncertainties or residuals; adding these would make the comparison between the fitted values and the source-reference Bragg vectors G_hi and G_lo quantitative.
[Fig. 5(a)] The Pareto plot would be easier to read if the axis labels and the meaning of the color or marker coding for different α values were explained in the caption.

Circularity Check

0 steps flagged · score 0.0 of 10

The paper shows no significant circularity: all load-bearing steps are supported by external Fourier identities, an external benchmark, and post-hoc fits that do not feed back into the optimization.

full rationale

The paper's central construction is self-contained: T_AC is defined from the weighted long-lag autocorrelation energy of the reflected field normalized by the incident pulse (Eq. 3), and its connection to spectral roughness is established by the external Wiener-Khinchin and Parseval identities rather than by assuming the conclusion. The constraint bound T_AC <= 1 is anchored to an external benchmark, explicitly 'in reference to a perfect reflector, for which eta_r = 1 and T_AC = 1', so it is not fitted to the optimized results. The statistical demonstration that constrained designs keep median reflectance 0.8096 vs 0.8093 while removing narrow features is an empirical outcome, not a forced one: nothing in the definition guarantees that suppressing the tail will preserve reflectance, and the unconstrained ensemble indeed shows T_AC values up to 8.628. Self-citations [9] and [13] supply only the generic adjoint and TopOpt methodology and are not used to justify the metric's validity. The chirp reconstruction in Sec. III-C fits parameters a1, a2, and D0 after optimization and then cross-checks them against Bragg estimates; because this fit does not feed back into the optimization or into the claimed constraint benefit, it is a post-hoc reverse-engineering exercise rather than a fitted-input prediction. The finite-FDTD-window concern raised by the skeptic is a numerical approximation issue, not a definitional reduction: the paper states that T_Sim is chosen long enough 'for the fields to decay', and even if incomplete decay biased some narrow features, that would be an accuracy limitation, not circularity in the derivation chain.

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

The central claim rests on standard Fourier analysis plus the numerical FDTD/adjoint machinery. All tunable parameters are heuristic; no new physical entities are introduced.

free parameters (3)

tau_min (1% Hilbert-envelope threshold) = 10.14 fs (for f0=300 THz, 2*Delta_f=105 THz pulse)
Chosen as the first lag where the Hilbert envelope of the incident pulse autocorrelation decays to 1% of its peak; defines the long-lag region in Eq. (3). The 1% threshold is ad hoc and affects the metric and the optimization.
Autocorrelation lag exponent (gamma) = 2
The quadratic weight (tau/tau_min)^2 in Eq. (3) is hand-picked to increase penalty on late tails; no derivation or cross-validation is provided.
Chirp fit coefficients a1, a2, D0 = a1=16.23 rad/um, a2=11.18 rad/um, D0=0.46
Fitted to the optimized design in Sec. III-C for the parametric reconstruction; post-hoc and not used in the optimization itself.

assumptions (4)

standard math Parseval-Plancherel theorem and Wiener-Khinchin theorem hold for the finite-energy signals considered.
Used in Sec. II to relate total energy and autocorrelation to spectral density and to justify E_AC as a spectral-derivative proxy.
domain assumption The 1D FDTD model with linear permittivity interpolation accurately represents the dielectric Bragg mirror physics (including reflection and dispersion).
Used throughout Sec. III and App. A; no experimental verification, only numerical.
domain assumption A perfect reflector has T_AC = 1 and eta_r = 1, justifying the constraint bound T_AC <= 1 in Eq. (6b).
Sec. III-A places the upper bound at 1 'in reference to a perfect reflector, for which eta_r = 1 and T_AC = 1'.
domain assumption The time-domain adjoint method as implemented yields the correct gradient of the (normalized) autocorrelation metric.
App. A states the adjoint source is the time-reversed derivative but does not derive it explicitly; correctness is assumed.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Spectrally smooth broadband response via autocorrelation-constrained inverse design." pith.science (2026). https://pith.science/paper/OX2POPYT

@misc{pith2026260813367,
  author       = {Pith},
  title        = {Pith review of: Spectrally smooth broadband response via autocorrelation-constrained inverse design},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OX2POPYT}},
  note         = {Machine review of arXiv:2608.13367}
}

read the original abstract

Time-domain inverse design in photonics is known to be suitable for maximizing the efficiency of optical devices over broad frequency ranges. Objectives commonly used in this context include time-integrated field quantities derived from Poynting's theorem, such as energy, flux, or dissipated power, which can be directly linked to integrated frequency-domain responses via Parseval's theorem. While computationally efficient, these objectives measure only the total response over the targeted bandwidth and, as we demonstrate, are insufficient to capture undesired in-band ripple, narrow spectral features, or sidelobes. We overcome this limitation by introducing a time-domain metric quantifying such spectral variations based on the weighted long-lag autocorrelation energy of the optical response. We incorporate this metric into an FDTD-based topology-optimization framework and demonstrate its beneficial effect on the example of inverse designing one-dimensional dielectric Bragg mirrors via the time-domain adjoint method.