Beyond the Parasitic Limit: A Nanoprobing Framework for De-embedding Intrinsic Ferroelectric Properties at the Deep Sub-Micrometer Scale

Adrian Petraru; Dirk Meyners; Fabian Lofink; Georg Sch\"onweger; Niklas Kyoushi; Ole Gronenberg; Simon Fichtner

REVIEW 3 major objections 7 minor 11 references

Beyond the Parasitic Limit: A Nanoprobing Framework for De-embedding Intrinsic Ferroelectric Properties at the Deep Sub-Micrometer Scale

T0 review · 3 major / 7 minor · reviewed 2026-07-12 · grok-4.5

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

Pith's one-line read A nanoprobing de-embedding method recovers bulk-like ferroelectric switching in AlScN capacitors down to 165 nm.

desk verdict Solid experimental metrology toolkit for deep-submicron AlScN; the Screened Power Law is useful but multi-parameter, so the perfectly flat εr is partly by construction. read the letter →

arxiv 2607.03588 v1 pith:7FOARALR submitted 2026-07-03 cond-mat.mtrl-sci

Niklas Kyoushi , Georg Schönweger , Ole Gronenberg , Dirk Meyners , Adrian Petraru , Fabian Lofink , Simon Fichtner This is my paper · ORCID

classification cond-mat.mtrl-sci

keywords AluminumScandiumNitrideAlScNFerroelectricityNanoprobingDeviceScalingDe-embeddingWurtziteFerroelectricsDielectricLoss

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

Scaling ferroelectrics for dense memory hits a measurement wall: at deep sub-micrometer sizes the true device signal is swamped by probe parasitics and edge fringing, so reported degradation may be instrumental rather than material. This paper supplies an in-situ nanoprobing protocol that contacts capacitors as small as 165 nm without bond pads, then uses a Screened Power Law model to subtract the near-field probe contribution from the measured capacitance. After that correction the relative permittivity of 20 nm AlScN stays size-independent near 17, the apparent rise in dielectric loss is shown to be geometric dilution, and leakage-compensated large-signal loops still display clear ferroelectric switching even when only a few dozen grains remain under the electrode. The result is a practical toolkit that lets device physicists decide whether further scaling is limited by the crystal itself or by how the measurement is performed.

What carries the argument

The Screened Power Law geometric factor Kf(r) = Kbase + A (d/r)^n exp(-r/L), which multiplies the classical Kirchhoff-Palmer edge capacitance so that the residual after subtraction of static and dynamic probe strays is the intrinsic parallel-plate term.

What would settle it

Measure an identical set of capacitors with a fully shielded coaxial nanoprobe or a calibrated on-chip de-embedding structure; if the extracted εr still varies systematically with radius after the Screened Power Law correction, the model is incomplete.

Watch

Extended reading notes

Core claim

After Screened-Power-Law de-embedding of probe parasitics and geometric fringing, 20 nm Al0.72Sc0.28N capacitors retain size-independent permittivity εr ≈ 17 and bulk-like ferroelectric switching (stable Ec and clear J-E peaks) down to 165 nm diameter (N < 30 grains).

Load-bearing premise

The Screened Power Law fully accounts for every near-field probe and knife-edge contribution, so whatever capacitance remains after subtraction is purely the intrinsic device response.

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

This is a careful methods paper that actually solves a real measurement problem: how to get quantitative C-V and J-E data on ferroelectric capacitors down to ~165 nm without bond pads. The new pieces are the Screened Power Law de-embedding of probe-cone and knife-edge fringing, the conductance-per-area analysis that recovers bulk tan δ after geometric dilution, and the leakage-compensated few-grain PUND loops. They span five orders of magnitude in area, use SEM-extracted effective radii, report RSS error bars, and show that the apparent drop in loss is a vacuum-shunt artifact. That package is useful for anyone working on wurtzite ferroelectrics or high-k nano-caps.

What they do well is the experimental hygiene. Dual-chip process, in-situ Pt cap, multi-sweep averaging with wake-up exclusion, FFT filtering, and explicit spreading-resistance modeling all look solid. The dilution equation quantitatively matches the observed tan-δ floor once Cstatic and Cdynamic are taken from the capacitance fit, and the recovered εr ≈ 17 and tan δ ≈ 1.41 % sit on literature values. The J-E peaks remain clear at N < 30 grains; that is real data, not just a model claim.

The soft spot is exactly the one the stress-test flags, and it is real but not fatal. After locking Cstatic from the r o0 floor and εr from the large pads, residual capacitance is forced into Kf(r) = Kbase + A (d/r)^n exp(-r/L). A, n and L (estimated ~15 µm) are free enough that a weak intrinsic size dependence or unmodeled multipole could be absorbed while still producing high R^{2} and a flat post-subtraction εr. The functional form is physically motivated (Kirchhoff-Palmer + Jackson edge singularity + SPM screening), and the exponent shift from sub- to super-linear is consistent with the tapered “foot,” but uniqueness is untested—no FEM cross-check, no alternative model comparison. So Fig. 2c is not an independent experimental result; it is the residual after a flexible geometric subtraction. That weakens the strongest wording of “size-independent permittivity down to 165 nm,” but does not erase the rest of the toolkit.

This is for people who measure or model sub-micron ferroelectrics and need a practical de-embedding recipe. It deserves a serious referee; the data and SI are thorough enough that the fit-uniqueness issue can be tightened in revision. I would cite the conductance scaling and the few-grain J-E protocol, and I would bring it to reading group if we are talking metrology or AlScN scaling.

Referee Report

3 major / 7 minor

Summary. The manuscript presents an in-situ SEM nanoprobing framework for electrical characterization of isolated Al0.72Sc0.28N capacitors from ~50 µm down to 165 nm diameter without lithographic bond pads. A three-component capacitance decomposition (bulk + static stray + dynamic fringing) is closed with a Screened Power Law geometric factor Kf(r) that is intended to capture near-field probe-cone coupling and knife-edge electrode singularities; after de-embedding, a size-independent permittivity εr≈17 is reported. Apparent nanoscale suppression of tanδ is attributed to lossless vacuum-shunt dilution and is recovered via area–perimeter conductance scaling to a bulk value ~1.41%. Large-signal PUND J–E loops with leakage subtraction and FFT filtering show clear switching peaks and Ec trends down to ~190 nm (few-grain limit). The work positions the protocol as a general metrological toolkit for deep-submicron wurtzite ferroelectrics.

Significance. If the de-embedding is robust, the paper removes a genuine measurement bottleneck for scaled ferroelectrics: direct electrical access to attofarad-level capacitors without pad parasitics, with quantitative recovery of permittivity, loss, and hysteresis in the N<30-grain regime. The experimental span (five orders of magnitude in area), SEM-segmented effective radii, RSS error propagation, multi-sweep averaging, and the independent conductance-scaling recovery of tanδ are concrete strengths. The quantitative match between the predicted loss floor (~0.71%) and the measured plateau (~0.76%) is a useful cross-check. The framework is of clear interest to the AlScN and nanoscale ferroelectric communities even if some model uniqueness questions remain.

major comments (3)

§2.3, Eqs. (1)–(3) and Table 1 / Fig. 2c: After locking Cstatic from the r→0 floor and εr from macroscopic Chip-1 pads, residual capacitance is forced into Cdynamic = Kf(r)·Cedge(r) with free parameters A, n, and (estimated) L in the Screened Power Law. This functional form is flexible enough to absorb weak intrinsic size dependence, higher-order multipoles, or contact-area variation while still yielding high R² and a flat post-subtraction εr. The flatness of Fig. 2c is therefore not fully independent of the geometric model. Please (i) report a sensitivity analysis (vary A, n, L within physically plausible bounds and show the residual εr(r)), (ii) test at least one alternative geometric form (e.g., pure power law without screening, or a fixed-Kf Kirchhoff–Palmer baseline), and (iii) state which parameters are locked versus free for each chip so that uniqueness of the size-independent εr
§3.1 and Supplementary Note 5: The screening length L is stated as “estimated at 15 µm” from SPM literature rather than constrained by the present capacitance-versus-radius data. Because L multiplies the exponential cutoff of the near-field term, its value directly affects how much of the nanoscale excess capacitance is attributed to probe geometry versus the device. Either fit L (with uncertainty) from the Chip-2 data or demonstrate that the extracted εr and Kf trends are stable over a documented range of L (e.g., 5–30 µm).
§3.3–3.4 and Fig. 4b / Fig. 5c: Ec scaling trends are interpreted as perimeter-leakage voltage-masking, but Chip 1 and Chip 2 come from distinct deposition runs with different top-Pt thicknesses (30 nm vs 20 nm). The manuscript already notes a batch offset in Ec. Please separate batch-to-batch structural variation from true dimensional scaling more cleanly (e.g., by reporting Ec only within each chip, or by quantifying thickness/interface differences) before attributing the progressive hardening primarily to perimeter shunt effects.

minor comments (7)

Abstract and §1: “apparent degradation of dielectric loss” is slightly ambiguous; the data show an apparent *suppression* (dilution) of tanδ, not degradation. Align wording with §3.2.
Fig. 1 caption: panel labels “c), d)” then “d)” again; renumber so that the 250 nm contact image is uniquely labeled.
Eq. (4) and surrounding text: tanδmeas is written with Cstatic in the numerator without an explicit tanδstatic≈0 factor; state the lossless assumption explicitly in the equation for clarity.
Table 1: “actual thickness … estimated at ≈17 nm to align with the permittivity previously reported” should be justified with an independent thickness metrology (XRR/TEM) rather than permittivity matching alone, or the circularity should be flagged.
§2.5 / Fig. S2: grain diameter “31.15” lacks units in the main text (nm is clear from SI but should appear in the sentence).
Several SI figure references (e.g., “Figure S 1”, “Figure S 3”) have inconsistent spacing; normalize labeling.
References: ensure consistent formatting of journal names and DOIs; a few entries (e.g., ASTM grain-size standard) are incomplete.

Circularity Check

2 steps flagged · score 4.0 of 10

Screened Power Law multi-parameter fit (Kf free A/n/Kbase after locking macroscopic εr) absorbs residuals so post-subtraction flat εr≈17 is partly enforced by construction; loss-floor 'prediction' re-uses the same Cstatic/Cdynamic; other observables remain independent.

fitted input called prediction [Sec. 2.3 Eqs. (1)–(3); Table 1; Sec. 3.1 Fig. 2c]
"By applying the Screened Power Law model (Eq. 3) to capture these effects, we successfully extracted a size-independent relative permittivity of εr ≈ 17 ± 3 across the entire scaling range (Figure 2c)… A fixed Intrinsic εr ≈ 16.94 extracted from the larger pads of Chip 1. The actual thickness of the AlScN film was estimated at ≈17 nm to align with the permittivity previously reported [19, 20]."

εr (and d) are locked from macroscopic pads to match literature; free parameters A, n, Kbase of Kf(r) = Kbase + A(d/r)^n exp(-r/L) are then fitted to the identical Cmeas(r) dataset so that Cmeas - Cstatic - Kf·Cedge recovers Cbulk ∝ r^{2} with that fixed εr. The reported flatness of deconvolved εr(r) is therefore the residual after a multi-parameter geometric subtraction whose uniqueness is untested; any weak intrinsic size dependence is absorbed into Kf while still yielding R^{2} > 0.9999.
fitted input called prediction [Sec. 3.2 Eq. (5) and surrounding text]
"Substituting the extracted values (Cstatic ≈ 70 fF, Cdynamic ≈ 72 fF) and the intrinsic bulk loss (tan δbulk ≈ 1.41 %) into the dilution equation: tan δfloor ≈ Cdynamic/(Cstatic + Cdynamic) · tan δbulk yields a theoretical floor of ≈0.71 %. This prediction is in excellent agreement with the experimental saturation level (≈0.76 %)."

Cstatic and Cdynamic that enter the 'prediction' of the loss floor are themselves outputs of the Screened Power Law capacitance fit performed on the same devices. The numerical agreement therefore largely reconfirms internal consistency of the earlier geometric decomposition rather than constituting an independent external validation of the dilution model.

full rationale

The central permittivity claim is not fully independent of its geometric model inputs. Cmeas(r) is decomposed as Cbulk(fixed εr) + Cstatic + Kf(r)·Cedge(r) with Kf a flexible three-parameter Screened Power Law; after locking εr (and adjusting film thickness) from large pads to literature values, free parameters are optimized on the same C-vs-r data so that residuals match parallel-plate scaling. High R^{2} therefore guarantees a flat deconvolved εr by construction of the subtraction; mild intrinsic size dependence or unmodeled multipoles can be absorbed into Kf without detection. The subsequent tanδ-floor calculation re-uses those fitted Cstatic/Cdynamic values, creating a secondary consistency check rather than a fully external validation. However, the functional form is physically motivated (Kirchhoff–Palmer + Jackson edge singularity + SPM screening), macroscopic references are external, and the conductance-area scaling plus leakage-compensated J-E peaks use separate observables and do not reduce to the same fit. Thus partial circularity exists around the εr claim but does not collapse the entire framework; score 4 reflects one load-bearing fitted-input step plus a secondary reuse, not definitional equivalence of the whole paper.

Assumptions & free parameters 5 free parameters · 4 assumptions · 1 invented entities

Central claim rests on classical electrostatic edge formulas, a phenomenological screening length taken from SPM literature, and several free parameters fitted to the measured C(r) curve. No new particles or forces are postulated; the Screened Power Law is an ad-hoc but physically motivated functional form.

free parameters (5)

Coupling amplitude A = 2107 / 5942
Fitted separately for wafer-prober (2107) and nanoprober (5942) chips; controls magnitude of near-field stray term.
Scaling exponent n = 0.86 / 1.12
Fitted; transitions from 0.86 (macro) to 1.12 (nano) and is interpreted as onset of conductive-fin singularity.
Screening length L = 15 µm
Estimated at 15 µm from SPM literature and held fixed; sets the exponential roll-off of dynamic stray capacitance.
Static stray Cstatic = 70-80 fF
Locked to asymptotic r o0 intercept (70-80 fF); subtracts the constant probe-arm background.
Base geometric factor Kbase = 1.0 / 1.5
Fitted (1.0 macro, 1.5 nano); absorbs constant sidewall-slope enhancement.

assumptions (4)

domain assumption Kirchhoff-Palmer circular-disk edge capacitance formula remains the correct baseline before geometric-factor scaling.
Invoked in §2.3 and SI Note 5 as the analytic starting point for Cedge(r).
domain assumption Jackson edge-singularity field divergence E∝ r^{v-1} with v=π/(2π-β) correctly describes the tapered 'knife-edge' electrode foot.
Used in SI Note 5 to justify super-linear n>1 at the nanoscale.
domain assumption Measured loss tangent is a capacitance-weighted average of lossy mesa and lossless vacuum contributions (mixture equation).
Eq. 4 and dilution calculation in SI Note 5; underpins the geometric-dilution claim.
standard math Spreading resistance of thin Pt electrodes is given by the 2-D radial integral of sheet resistivity.
SI Note 6; used to show contact resistance is negligible compared with measured ESR.

invented entities (1)

Screened Power Law geometric factor Kf(r)
purpose: Provides a continuous, radius-dependent multiplier that bridges plate, transition and tip/cone electrostatic regimes so that intrinsic Cbulk can be isolated.
Functional form is introduced ad hoc in Eq. 3; parameters are fitted to data rather than derived from first principles.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Beyond the Parasitic Limit: A Nanoprobing Framework for De-embedding Intrinsic Ferroelectric Properties at the Deep Sub-Micrometer Scale." pith.science (2026). https://pith.science/paper/7FOARALR

@misc{pith2026260703588,
  author       = {Pith},
  title        = {Pith review of: Beyond the Parasitic Limit: A Nanoprobing Framework for De-embedding Intrinsic Ferroelectric Properties at the Deep Sub-Micrometer Scale},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7FOARALR}},
  note         = {Machine review of arXiv:2607.03588}
}

read the original abstract

The continued scaling of ferroelectric devices is critical for next-generation computing architectures, yet it is fundamentally challenged by a metrological bottleneck: at the deep sub-micrometer scale, intrinsic material properties are heavily masked by extrinsic parasitic impedances and geometric fringing fields. Here, we introduce a quantitative, in-situ nanoprobing framework capable of resolving the true electrical response of ferroelectric capacitors down to 165 nm in diameter without the need for lithographic bond pads. Using 20 nm thick AlScN as a model system, we establish a non-linear 'Screened Power Law' model to decouple attofarad-level device capacitances from massive near-field probe interactions. Furthermore, we demonstrate that the apparent degradation of dielectric loss at the nanoscale is a geometric dilution artifact, which we overcome through a conductance scaling analysis. Finally, we apply this framework to large-signal characterization, utilizing leakage-compensation and noise filtering protocols to extract pristine intrinsic hysteresis (C-V and J-E) loops in the discrete few-grain limit. These findings provide a universal analytical toolkit required to overcome the measurement limits of deep-submicron ferroelectrics.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

11 extracted references

[1]

Lossless Static Stray (𝐶static): Attributed to the probe arm and manipulator body coupling through the vacuum/air dielectric, modeled as effectively lossless (tan𝛿meas ≈ 0)
[2]

lossless dilution

Lossy Dynamic Stray ( 𝐶dynamic): Attributed to the active tip cone interacting with the ferroelectric AlScN mesa, which carries the material's intrinsic loss (tan𝛿AlScN). Consequently, the measured loss tangent tan𝛿meas describes a dilution of the intrinsic material loss by the lossless parasitic background, defined by the mixture equation: tan𝛿meas(𝑟) ≈ ...

2019
[3]

Microscale Fabrication (Chip 1, 50 µm – 5 µm): Devices in the 50 µm to 5 µm range were patterned using standard UV lithography (Karl Süss MA6, AZ 1518, MicroChemicals GmbH). The top Pt electrode was structured via Ion Beam Etching (IBE), timed to stop with a controlled over-etch at the AlScN surface verified by secondary -ion mass spectroscopy (SIMS) thro...
[4]

In this step, the etch proceeded through the AlScN layer and stopped at the bottom Pt electrode, creating isolated islands and exposing the bottom contact

Nanoscale Fabrication (Chip 2, 2 µm – 100 nm): Deep-sub micrometer devices required a two -step hybrid process to ensure electrical isolation and precise definition: • Step A ( Mesa Definition): Large-area mesas (20×20µm²) were defined using UV lithography and IBE. In this step, the etch proceeded through the AlScN layer and stopped at the bottom Pt elect...
[5]

They feature a shaft diameter ranging from 80 nm to 100 nm (measured at a height of 50 nm above the apex)

High-Resolution Probes: Mesoscope probes were selected for the smallest devices and to measure the small-signal capacitance and dielectric loss measurement series in the nanoprober. They feature a shaft diameter ranging from 80 nm to 100 nm (measured at a height of 50 nm above the apex)
[6]

Robust Probes: For larger devices and general contact, Omniprobe AutoProbe tips were used, characterized by a tip radius of 0.5 µm (corresponding to a diameter of ~200 nm at 50 nm above the apex) and a taper angle of 10°-13°. 17
[7]

Conductive Fin

Waferprober Probes: AmericanProbes tungsten wire probes on a nickel shank were utilized with a tip radius of 0.35 µm. All probes used in the nanoprober setup were inserted into adapters and mechanically bent downwards by approximately 40° at a distance of 2-3 mm from the tip. This geometrical adaptation was essential to align the probe’s actuation axes wi...
[8]

Due to finite pixel resolution and the polygonal faceting inherent to electron-beam lithography at the deep sub-micrometer scale, a radial uncertainty of ±5 % was assigned

Geometric Uncertainty ((∆𝑨): The dimensional uncertainty is dictated by the spatial resolution limits of the scanning electron microscopy (SEM) utilized to extract the effective electrode radius ( 𝑟eff). Due to finite pixel resolution and the polygonal faceting inherent to electron-beam lithography at the deep sub-micrometer scale, a radial uncertainty of...

Show all 11 references

[9]

Systematic Instrumental Accuracy (∆𝑪𝐬𝐲𝐬): In the sub -micrometer limit, the total measured parallel impedance (|𝑍m| = [(1 𝑅p⁄ ) 2 + (𝜔𝐶p) 2 ] −1 2⁄ ) increases exponentially, severely degrading the performance of standard bridge -balance metrology. The systematic capacitance e...
[10]

Initial analysis of the raw, single-sweep traces for the deep sub-micrometer devices revealed a Signal-to-Noise Ratio (SNR) approaching unity

Statistical Variance and SNR (∆𝑪𝐬𝐭𝐚𝐭): Beyond systematic limits, random thermal and environmental noise (∆𝐶stat) masks the intrinsic displacement current at the nanoscale. Initial analysis of the raw, single-sweep traces for the deep sub-micrometer devices revealed a Signal-to...
[11]

butterfly

Total Propagated Uncertainty To establish the absolute metrological limits of the extraction framework, the total absolute capacitance error was calculated by adding the systematic instrument limits and the statistical variance in quadrature: ∆𝐶total = √𝐶sys2 + 𝐶stat 2 (15) As...

2016

Pith tools

Reviewed July 12, 2026 · model on record in the stance chip above.

[1] [1]

Lossless Static Stray (𝐶static): Attributed to the probe arm and manipulator body coupling through the vacuum/air dielectric, modeled as effectively lossless (tan𝛿meas ≈ 0)

[2] [2]

lossless dilution

Lossy Dynamic Stray ( 𝐶dynamic): Attributed to the active tip cone interacting with the ferroelectric AlScN mesa, which carries the material's intrinsic loss (tan𝛿AlScN). Consequently, the measured loss tangent tan𝛿meas describes a dilution of the intrinsic material loss by the lossless parasitic background, defined by the mixture equation: tan𝛿meas(𝑟) ≈ ...

2019

[3] [3]

Microscale Fabrication (Chip 1, 50 µm – 5 µm): Devices in the 50 µm to 5 µm range were patterned using standard UV lithography (Karl Süss MA6, AZ 1518, MicroChemicals GmbH). The top Pt electrode was structured via Ion Beam Etching (IBE), timed to stop with a controlled over-etch at the AlScN surface verified by secondary -ion mass spectroscopy (SIMS) thro...

[4] [4]

In this step, the etch proceeded through the AlScN layer and stopped at the bottom Pt electrode, creating isolated islands and exposing the bottom contact

Nanoscale Fabrication (Chip 2, 2 µm – 100 nm): Deep-sub micrometer devices required a two -step hybrid process to ensure electrical isolation and precise definition: • Step A ( Mesa Definition): Large-area mesas (20×20µm²) were defined using UV lithography and IBE. In this step, the etch proceeded through the AlScN layer and stopped at the bottom Pt elect...

[5] [5]

They feature a shaft diameter ranging from 80 nm to 100 nm (measured at a height of 50 nm above the apex)

High-Resolution Probes: Mesoscope probes were selected for the smallest devices and to measure the small-signal capacitance and dielectric loss measurement series in the nanoprober. They feature a shaft diameter ranging from 80 nm to 100 nm (measured at a height of 50 nm above the apex)

[6] [6]

Robust Probes: For larger devices and general contact, Omniprobe AutoProbe tips were used, characterized by a tip radius of 0.5 µm (corresponding to a diameter of ~200 nm at 50 nm above the apex) and a taper angle of 10°-13°. 17

[7] [7]

Conductive Fin

Waferprober Probes: AmericanProbes tungsten wire probes on a nickel shank were utilized with a tip radius of 0.35 µm. All probes used in the nanoprober setup were inserted into adapters and mechanically bent downwards by approximately 40° at a distance of 2-3 mm from the tip. This geometrical adaptation was essential to align the probe’s actuation axes wi...

[8] [8]

Due to finite pixel resolution and the polygonal faceting inherent to electron-beam lithography at the deep sub-micrometer scale, a radial uncertainty of ±5 % was assigned

Geometric Uncertainty ((∆𝑨): The dimensional uncertainty is dictated by the spatial resolution limits of the scanning electron microscopy (SEM) utilized to extract the effective electrode radius ( 𝑟eff). Due to finite pixel resolution and the polygonal faceting inherent to electron-beam lithography at the deep sub-micrometer scale, a radial uncertainty of...