arxiv: 2605.03778 · v1 · submitted 2026-05-05 · ⚛️ physics.optics · cond-mat.mtrl-sci

Recognition: unknown

Neural-network reconstruction of THz transmission spectra using electrically tunable AlGaN/GaN plasmonic-crystal analyzer

A. Witkowska , M. Dub , P. Sai , P. Tiwari , M. Sakowicz , J. A. Majewski , W. Knap

Authors on Pith no claims yet

Pith reviewed 2026-05-07 14:32 UTC · model grok-4.3

classification ⚛️ physics.optics cond-mat.mtrl-sci

keywords terahertz spectroscopyplasmonic crystalneural network reconstructionAlGaN/GaNspectral inversionvoltage-tunable filterinterferometer-freeTHz transmission

0 comments

The pith

Electrically tunable plasmonic-crystal analyzer combined with neural-network inversion reconstructs THz spectra without an interferometer.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper shows that an electrically tunable grating-gate AlGaN/GaN plasmonic crystal encodes a THz transmission spectrum as voltage-dependent intensity readings. A feedforward neural network trained on synthetic data inverts those readings to recover the spectrum. Validation on four real samples in both FTIR and direct fixed-mirror modes yields low mean squared error and identifies six of seven ground-truth resonances while outperforming Tikhonov regularization with fewer spurious peaks. The work removes the need for mechanical interferometers in THz spectroscopy. A reader would care if this leads to simpler, more compact spectral tools.

Core claim

Voltage-tunable plasmonic filtering combined with neural-network inversion establishes an interferometer-free architecture for THz spectral reconstruction. The network achieves a mean square error of the reconstruction of 0.015 in FTIR mode and 0.038 in direct mode, correctly identifying six out of seven ground-truth resonances in each mode. Against a first-difference Tikhonov regularization baseline, the mean reconstruction error is reduced 3.6 times in FTIR mode and 1.55 times in direct mode, with fewer spurious peaks and lower peak-position errors.

What carries the argument

Electrically tunable grating-gate AlGaN/GaN plasmonic-crystal analyzer that encodes spectral content into voltage-dependent intensity signals inverted by a feedforward neural network.

If this is right

The approach supports spectral reconstruction in direct fixed-mirror acquisition without mechanical moving parts.
Mean reconstruction error drops relative to Tikhonov regularization baselines, with reduced spurious peaks.
Six of seven ground-truth resonances are recovered correctly on the validated samples in both operating modes.
The architecture eliminates the interferometer while maintaining usable accuracy for THz transmission spectra.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

Compact voltage-controlled devices could enable portable or on-chip THz spectrometers for real-time material or gas sensing.
The same encoding-plus-inversion pattern may extend to other tunable filters or spectral ranges where mechanical interferometers are impractical.
Collecting more diverse experimental training data could further reduce domain-shift errors when deploying the network on unseen samples.

Load-bearing premise

The synthetic dataset used to train the neural network accurately represents the real experimental conditions, noise, and device response of the four tested samples, allowing reliable generalization without significant domain shift.

What would settle it

Testing the trained network on new samples whose resonances or noise levels fall outside the synthetic training distribution and finding substantially higher reconstruction error or missed resonances would show the method does not generalize as claimed.

Figures

Figures reproduced from arXiv: 2605.03778 by A. Witkowska, J. A. Majewski, M. Dub, M. Sakowicz, P. Sai, P. Tiwari, W. Knap.

**Figure 1.** Figure 1: (a) FTIR measured transmittance of the S7 analyzer for various applied gate view at source ↗

**Figure 2.** Figure 2: Synthetic test-set reconstruction. Top row: one random example per generator view at source ↗

**Figure 3.** Figure 3: Comparison of the signal in FTIR and direct acquisition modes for the reference view at source ↗

**Figure 4.** Figure 4: First-difference Tikhonov reconstruction view at source ↗

read the original abstract

We demonstrate machine learning (ML) based reconstruction of terahertz transmission spectra using an electrically tunable grating-gate AlGaN/GaN plasmonic-crystal analyzer. The analyzer encodes the transmission spectrum into a voltage-dependent intensity, which is then inverted by an ML algorithm. A feedforward neural network trained on a synthetic dataset is validated experimentally on four samples in standard Fourier Transform Infrared (FTIR) mode and in direct (fixed-mirror) acquisition mode. The network achieves a mean square error (MSE) of the reconstruction of 0.015 in FTIR mode and 0.038 in direct mode, correctly identifying six out of seven ground-truth resonances in each mode. Against a first-difference Tikhonov regularization baseline, the mean reconstruction error is reduced 3.6 times in FTIR mode and 1.55 times in direct mode, with fewer spurious peaks and lower peak-position errors. Voltage-tunable plasmonic filtering combined with neural-network inversion establishes an interferometer-free architecture for THz spectral reconstruction.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper shows a working experimental demo of voltage-tunable GaN plasmonic filtering plus neural-net inversion for THz spectra, with concrete gains over Tikhonov on four real samples.

read the letter

The main point is that they encode a THz transmission spectrum into a set of voltage-dependent intensity readings from an electrically tunable AlGaN/GaN grating-gate plasmonic crystal, then train a feedforward network on synthetic data to recover the spectrum. On four tested devices this gives MSE of 0.015 in FTIR mode and 0.038 in direct mode, correctly flags six of seven ground-truth resonances, and cuts reconstruction error by 3.6x and 1.55x versus a first-difference Tikhonov baseline while producing fewer spurious peaks. The experimental validation and the direct comparison to a standard regularizer are the parts that land. The hardware itself is a reasonable choice for a compact, no-moving-parts analyzer because the plasmon resonances shift with gate voltage and give multiple independent measurements from one device. The soft spot is the synthetic training set. The network never sees real noise statistics, contact effects, or fabrication scatter during training, and the reported numbers come from exactly the same four samples used for validation. Without an ablation on held-out real measurements or a clear check that the generator reproduces the measured transmission curves and voltage shifts, it is hard to know how much domain shift is hidden in the good results. The abstract also leaves the network depth, training details, and exact synthetic generator unspecified. This is aimed at people working on compact THz spectrometers or plasmonic sensors who need a practical reconstruction step. A reader who already follows ML inversion in spectroscopy will find the numbers and baseline comparison useful. I would send it to peer review. The experimental results and the improvement over the baseline give it enough substance to justify referee time, even though the synthetic-data validation will need tightening.

Referee Report

1 major / 2 minor

Summary. The paper demonstrates machine learning-based reconstruction of THz transmission spectra using an electrically tunable grating-gate AlGaN/GaN plasmonic-crystal analyzer. The device encodes the spectrum into voltage-dependent intensity measurements, which a feedforward neural network (trained on synthetic data) inverts to recover the spectrum. Experimental validation on four samples yields MSE of 0.015 (FTIR mode) and 0.038 (direct mode), correctly identifies 6/7 ground-truth resonances, and reduces mean reconstruction error by factors of 3.6 and 1.55 relative to a first-difference Tikhonov baseline, with fewer spurious peaks.

Significance. If the synthetic training data faithfully captures real-device response, noise, and voltage tuning, the work establishes a compact interferometer-free architecture for THz spectroscopy. The concrete MSE values, resonance identification rates, and quantitative baseline comparison on multiple experimental samples constitute a clear strength; the approach could enable simpler THz instrumentation if the domain-shift concern is resolved.

major comments (1)

[validation results and methods sections] The central claim that the NN trained solely on synthetic spectra inverts real experimental data rests on unverified fidelity of the synthetic generator to measured transmission curves, plasmon resonance shifts with gate voltage, detector noise, and fabrication variations across the four tested devices. No ablation, noise-injection test, or cross-validation against held-out real measurements is described that would quantify any domain gap (see validation results and methods sections).

minor comments (2)

[abstract and results] The abstract and results would benefit from explicit statement of the NN architecture (layer count, activation functions, training hyperparameters) and the precise procedure used to generate the synthetic dataset.
[figures and tables] Figure captions and table legends should clarify whether the reported resonance identification success and MSE values are averaged over the four samples or reported per sample.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the constructive feedback and the opportunity to clarify the validation strategy. We address the concern about synthetic-data fidelity below and outline planned revisions.

read point-by-point responses

Referee: The central claim that the NN trained solely on synthetic spectra inverts real experimental data rests on unverified fidelity of the synthetic generator to measured transmission curves, plasmon resonance shifts with gate voltage, detector noise, and fabrication variations across the four tested devices. No ablation, noise-injection test, or cross-validation against held-out real measurements is described that would quantify any domain gap (see validation results and methods sections).

Authors: We agree that explicit quantification of the domain gap would strengthen the manuscript. The synthetic generator is constructed from a physics-based model of the grating-gate plasmonic response (Drude-Lorentz permittivity with voltage-tuned carrier density) whose resonance positions and linewidths are fitted to measured device characteristics; additive Gaussian noise is calibrated to the experimental detector SNR. Performance on four independent real samples (MSE 0.015/0.038, 6/7 resonances recovered) provides an empirical test of generalization. However, the current text does not contain ablations, noise-injection sweeps, or direct synthetic-vs-measured curve overlays. In the revised manuscript we will (i) expand the Methods section with the full synthetic-data equations and parameter-extraction procedure, (ii) add a figure comparing synthetic and experimental transmission spectra for one device, and (iii) include a noise-injection ablation that reports reconstruction MSE versus added noise level. These additions will directly quantify the domain gap. revision: yes

Circularity Check

0 steps flagged

No circularity: reconstruction validated directly against experimental ground-truth spectra

full rationale

The paper trains a feedforward neural network on synthetic spectra generated from a model of the voltage-tunable plasmonic analyzer and evaluates reconstruction quality by direct comparison to measured FTIR and direct-mode transmission spectra from four physical samples. Reported MSE values (0.015 FTIR, 0.038 direct) and resonance identification counts are computed against these independent experimental references, not against the synthetic generator itself. No equations, parameters, or self-citations are shown that would make the output equivalent to the training inputs by construction; the Tikhonov baseline comparison is also external. The derivation chain therefore remains self-contained and falsifiable outside the fitted network weights.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on standard assumptions in machine learning generalization from synthetic to real data and on the deterministic encoding properties of the plasmonic device; no new free parameters or invented entities are introduced beyond routine neural network training.

axioms (1)

domain assumption The voltage-dependent intensity measurements encode sufficient spectral information that a neural network trained on synthetic data can invert to recover the original transmission spectrum on experimental samples.
Implicit in the encoding-inversion pipeline and the reported experimental validation.

pith-pipeline@v0.9.0 · 5504 in / 1229 out tokens · 125634 ms · 2026-05-07T14:32:39.126555+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

25 extracted references · 1 canonical work pages · 1 internal anchor

[1]

Identification and quantification of polymorphism in the pharmaceutical compound diclofenac acid by terahertz spectroscopy and solid-state density functional theory,

M. D. King, W. D. Buchanan, and T. M. Korter, “Identification and quantification of polymorphism in the pharmaceutical compound diclofenac acid by terahertz spectroscopy and solid-state density functional theory,” Anal. Chem.83, 3786–3792 (2011)

2011
[2]

Using terahertz pulsed spectroscopy to quantify pharmaceutical polymorphism and crystallinity,

C. J. Strachan, P. F. Taday, D. A. Newnham,et al., “Using terahertz pulsed spectroscopy to quantify pharmaceutical polymorphism and crystallinity,” J. Pharm. Sci.94, 837–846 (2005)

2005
[3]

Application of terahertz detection technology in non-destructive thickness measurement,

H. Liet al., “Application of terahertz detection technology in non-destructive thickness measurement,” Photonics12, 1191 (2025)

2025
[4]

Terahertz imaging and spectroscopy in cancer diagnostics: A technical review,

Y. Penget al., “Terahertz imaging and spectroscopy in cancer diagnostics: A technical review,” BME Front.2020, 2547609 (2020)

2020
[5]

Nanoscalebiomaterialsforterahertzimaging: Anon-invasiveapproachforearlycancerdetection,

A.Sadeghietal.,“Nanoscalebiomaterialsforterahertzimaging: Anon-invasiveapproachforearlycancerdetection,” Transl. Oncol.27, 101565 (2023)

2023
[6]

Enhancing concealed object detection in active THz security images with adaptation-YOLO,

A. Chenget al., “Enhancing concealed object detection in active THz security images with adaptation-YOLO,” Sci. Rep.15, 2735 (2025)

2025
[7]

Recent developments in spectroscopic techniques for the detection of explosives,

W. Zhanget al., “Recent developments in spectroscopic techniques for the detection of explosives,” Materials11, 1364 (2018)

2018
[8]

Tutorial: An introduction to terahertz time domain spectroscopy (THz-TDS),

J. Neu and C. A. Schmuttenmaer, “Tutorial: An introduction to terahertz time domain spectroscopy (THz-TDS),” J. Appl. Phys.124, 231101 (2018)

2018
[9]

Frequency-domain terahertz spectroscopy using long-carrier-lifetime photoconductive antennas,

P.-K. Lu and M. Jarrahi, “Frequency-domain terahertz spectroscopy using long-carrier-lifetime photoconductive antennas,” Opt. Express31, 9319 (2023)

2023
[10]

Bründermann, H.-W

E. Bründermann, H.-W. Hübers, and M. F. Kimmitt,Terahertz Techniques, vol. 151 ofSpringer Series in Optical Sciences(Springer, Berlin, Heidelberg, 2012)

2012
[11]

Far infrared spectroscopy with high resolution cyclotron resonance filters,

C. Skierbiszewskiet al., “Far infrared spectroscopy with high resolution cyclotron resonance filters,” J. Appl. Phys. 84, 433 (1998)

1998
[12]

Electrically tunable topological notch filter for THz integrated photonics,

M. Guptaet al., “Electrically tunable topological notch filter for THz integrated photonics,” Adv. Opt. Mater.11, 2301051 (2023)

2023
[13]

Terahertz notch and low-pass filters based on band gaps properties by using metal slits in tapered parallel-plate waveguides,

E. S. Leeet al., “Terahertz notch and low-pass filters based on band gaps properties by using metal slits in tapered parallel-plate waveguides,” Opt. Express19, 14852 (2011)

2011
[14]

Electrical tuning of terahertz plasmonic crystal phases,

P. Saiet al., “Electrical tuning of terahertz plasmonic crystal phases,” Phys. Rev. X13, 041003 (2023)

2023
[15]

P. C. Hansen,Discrete Inverse Problems: Insight and Algorithms(SIAM, Philadelphia, 2010)

2010
[16]

Experimental demonstration of infrared spectral reconstruction using plasmonic metasurfaces,

B. Craig, V. R. Shrestha, J. Meng,et al., “Experimental demonstration of infrared spectral reconstruction using plasmonic metasurfaces,” Opt. Lett.43, 4481 (2018)

2018
[17]

Single-shot on-chip spectral sensors based on photonic crystal slabs,

Z. Wanget al., “Single-shot on-chip spectral sensors based on photonic crystal slabs,” Nat. Commun.10, 1020 (2019)

2019
[18]

Neural network-based on-chip spectroscopy using a scalable plasmonic encoder,

C. Brownet al., “Neural network-based on-chip spectroscopy using a scalable plasmonic encoder,” ACS Nano15, 6305 (2021)

2021
[19]

Deeply learned broadband encoding stochastic hyperspectral imaging,

W. Zhanget al., “Deeply learned broadband encoding stochastic hyperspectral imaging,” Light Sci. Appl.10, 108 (2021)

2021
[20]

Imaging-based intelligent spectrometer on a plasmonic rainbow chip,

D. Tuaet al., “Imaging-based intelligent spectrometer on a plasmonic rainbow chip,” Nat. Commun.14, 1902 (2023)

1902
[21]

Miniaturization of optical spectrometers,

Z. Yang, T. Albrow-Owen, W. Cai, and T. Hasan, “Miniaturization of optical spectrometers,” Science371, eabe0722 (2021)

2021
[22]

Tunable THz spectrum analyzer with hyperspectral resolution,

X. Heet al., “Tunable THz spectrum analyzer with hyperspectral resolution,” Chin. Opt. Lett.22, 073601 (2024)

2024
[23]

Ultracompact terahertz spectrometer based on magneto-electric-coupled optoelectronic metasurface assisted by deep learning,

S. Zhao, Y. Wen, C. Wang,et al., “Ultracompact terahertz spectrometer based on magneto-electric-coupled optoelectronic metasurface assisted by deep learning,” Laser Photonics Rev.20, e01565 (2026)

2026
[24]

Searching for Activation Functions

P. Ramachandran, B. Zoph, and Q. V. Le, “Searching for activation functions,” arXiv preprint arXiv:1710.05941 (2017)

work page internal anchor Pith review arXiv 2017
[25]

Decoupled weight decay regularization,

I. Loshchilov and F. Hutter, “Decoupled weight decay regularization,” in7th International Conference on Learning Representations (ICLR),(2019)

2019