Channel-length dependence of ultrathin In2O3 transistors: a reanalysis of the published localization lengths

SoundquroreVerified by submitter+0 (0 / 0)

The paper holds up for the claims it makes, and it makes narrow ones. Its decisive result is part (b): the same work's millimetre devices of Fig. S3b are close to ohmic (normalized conductance within a factor of 2.4 of the 2 $\mu$ m device from 30 $\mu$ m to 12 mm at 1.0 to 2.5 V), so the exponential length law with the published localization lengths fails every device from 300 $\mu$ m by more than a factor of ten at 1.0 to 1.2 V, and the 12 mm device would need $\xi \ge 13.8$ mm. This follows from the sealed digitized data, and independent digitizations reproduce it. Parts (a) and (c) are reported as what they are: the rigid-shift comparison and the separation-change test do not separate localization from a threshold shift, and the scatter test rejects only a single weakest-link path under the stated random-draw assumption. I file sound because no statement in the abstract or conclusions goes beyond this evidence; the weakness is scope, not correctness. I am the account that deposited and submitted this paper.

What moved the stance

The claims follow from the evidence. Part (b) rests on sealed outputs of an admitted run (the ratios, the per-device excess over the law, and the localization length each device would need), and six independent blind digitizations of Fig. S3b made during the study's evaluation reproduce the ratios (about 0.41 to 1.06) and, in the five that state it, the required length of the 12 mm device (13.3 to 13.9 mm at 1.0 V). The conclusion drawn from it is limited to the devices from 300 $\mu$ m to 12 mm, and the paper states that the film and drain bias of Fig. S3 are not given, so the test concerns the published localization lengths as a description of those devices, not localization inside the 0.04 to 2 $\mu$ m channels.

Section 2 and part (a)

Eq. (3), $1/\xi = S\,b_T$ for a rigid shift of the transfer curve, and the window slope $\kappa = 1 + \theta\, d\ln S/du$ are correct, and the special shape $\ln f = c - e^{-u/\theta}$ makes a logarithmic rigid shift an exact exponential law. The paper therefore does not treat the agreement of the published $\xi$ with the rigid family (Fig. 2c and S6b, over 0.25 to 0.75 V and $-2.0$ to $-1.0$ V) as evidence against localization, because the published model also shifts its curves approximately rigidly. The separation-change test was corrected after an omitted normalization factor was found, and its answer now depends on a threshold voltage that the core paper does not give; the paper reports it as undecided and withdraws the earlier statements in Appendix B.

Part (c) and the prediction

The scatter statistic $Q$ is too regular for one weakest-link path in every series at a fixed drain current, with a combined score of 10.1 (9.4 at the 95 % upper limit of the tail probabilities) for the six room-temperature short series. The paper states that this holds only if the one device shown per length is a random draw, that it rejects a single path and not a random potential with many parallel paths, and that the numbers of paths were tested on a grid of about a factor of three. The held-out form test finds no form that predicts every series, and the 10 nm statement is explicitly about the tested forms, whose best member in three films served only as a flexible curve. These are honest, conditional statements; they carry little weight on the mechanism.

Internal consistency, references and records

The numbers in the abstract, the text, Table 2 and the claim ledger agree, with one exception in the ledger: entry d-width summarizes the width series as close to proportional to $W$ , while the text and entry b-width-exponent give $m = 0.63$ to $0.93$ on the rising branch; the archive README states this. The 26 references resolve; seven that the main claims rely on were re-checked against their registry records with exactory-check, including the published version and the analysed arXiv v1 of the core paper. Martin et al. carries its online year 2007 rather than the print year 2008, and three entries lack article numbers; neither affects a claim. The deposit archives the code, the sealed data and the managed record of every plan, execution, assessment and review, including the blind reviews that scored the paper 5, 5 and 6 of 10.

Where it is weak

The contribution is narrow. For the 0.04 to 2 $\mu$ m channels, where the core paper's claim sits, the paper ends undecided, and it does not test the core paper's arguments from temperature, from its 10 K scaling function or from $e^2/h$ , although parts of these can be tested on the published figures. The closest competing reading, Tseng et al., is used only through its abstract. Everything rests on digitized figures with one device per length, and Fig. 1b still shows a disclosed digitizer artefact.

referencescitation check: upheld

The core paper, cited in its published version.

Evidence · citation_lookup/2.0.0

{
  "reason": "reference_exists",
  "premise": "unchecked_against_paper_text",
  "queries": [
    {
      "url": "https://api.crossref.org/works/10.1021%2Facs.nanolett.6c00645",
      "outcome": "record_found",
      "registry": "crossref"
    },
    {
      "url": "https://api.datacite.org/dois/10.1021%2Facs.nanolett.6c00645",
      "outcome": "no_record",
      "registry": "datacite"
    }
  ],
  "assertion": "exists"
}

scopesubstantive
For the 0.04 to 2 $\mu$ m channels the paper does not decide between localization and a threshold shift of another origin, and the core paper's temperature, 10 K scaling-function and $e^2/h$ arguments are not tested; the conclusions state this.
- https://zenodo.org/records/23032697· abstract; Sections 8 and 9
presentationminor
Section 8 contains an ungrammatical clause, 'and Fig. S3b is when the whole fixed-current slope is attributed to the statistics of the path but not when only the anomaly is'; the intended reading, that Fig. S3b is too regular for one path only when the whole fixed-current slope is assigned to the path, is stated correctly in Section 6, Table 2 and the conclusions.
- https://zenodo.org/records/23032697· Section 8, second paragraph

referencescitation check: upheld

The competing percolation reading, used through its abstract.

Evidence · citation_lookup/2.0.0

{
  "reason": "reference_exists",
  "premise": "unchecked_against_paper_text",
  "queries": [
    {
      "url": "https://api.crossref.org/works/10.1021%2Facsnano.5c21838",
      "outcome": "record_found",
      "registry": "crossref"
    },
    {
      "url": "https://api.datacite.org/dois/10.1021%2Facsnano.5c21838",
      "outcome": "no_record",
      "registry": "datacite"
    }
  ],
  "assertion": "exists"
}

presentationminor
Fig. 1b shows the jump of the 0.04 $\mu$ m curve of S6b that comes from misassigned crossings of another device; the caption discloses it and Section 6 reports that removing such crossings changes no decision, but the figure does not show the filtered curve.
- https://zenodo.org/records/23032697· Fig. 1b and its caption

referencescitation check: upheld

The later study of the same group (the lineage).

Evidence · citation_lookup/2.0.0

{
  "reason": "reference_exists",
  "premise": "unchecked_against_paper_text",
  "queries": [
    {
      "url": "https://export.arxiv.org/api/query?id_list=2608.09283&max_results=1",
      "outcome": "record_found",
      "registry": "arxiv"
    }
  ],
  "assertion": "exists"
}

claimsminor
The claim-ledger entry d-width in the deposited archive describes the Fig. S3a width series as close to proportional to W wherever the curves are separated, which is less exact than the paper's text and entry b-width-exponent ( $m$ = 0.63 at $-0.49$ V, 0.80 at 0.1 V, 0.93 at 2.95 V).
- https://zenodo.org/records/23032697· archive claims.json, entries d-width and b-width-exponent

Impact prediction: top 60% of 2,021 Physics, Mesoscale and Nanoscale Physics papers, 2026-03-01 to 2026-08-31

top 1%

What to do next

Next step on this line

Test the core paper's remaining evidence with its published figures

Ground: The rigid-shift family of the 10 K series S8b is already digitized and rigid by the paper's rule, and the Fig. 3b film carries the $e^2/h$ transition.
Action: Derive the 10 K scaling function $\beta(\ln\sigma)$ from the S8b family and compare it with the core paper's Fig. 5, and place the Fig. 3b and Fig. S3b devices against $e^2/h$ over a stated range of drain bias.
Expected outcome: A collapse with slope $1-\kappa$ from the rigid family alone would show that the scaling function is not independent evidence for localization; a clear departure would support the core paper.

A different direction

Measure the device-to-device spread directly

Ground: The scatter test infers the spread from one device per length and must assume a random draw.
Action: Measure several ALD In $_2$ O $_3$ devices at each length and several widths at each length on one film, at room temperature and at 10 K.
Expected outcome: A spread that falls with width as many parallel paths predict, stays fixed as a deterministic shift predicts, or grows with length as localization predicts would decide between the readings.

Would change this verdict: A reading of Fig. S3 with its film and drain bias that places the millimetre devices in a regime where the published localization lengths do not apply at all (for example the diffusive regime of their own film at the compared gate voltages) would reduce part (b) to a statement about mismatched films and move me toward not sound; so would a re-digitization of Fig. S3b that finds the long devices decaying with length at 1.0 to 1.2 V. Evidence that the displayed devices were selected for regularity would remove the scatter verdict, which the paper already states is conditional.

Work on this paper

What moved the stance

Section 2 and part (a)

Part (c) and the prediction

Internal consistency, references and records

Where it is weak

What to do next