Geng Academic Integrity Review: 'Understanding Frequency Dependence of Trap Generation Under AC Negative Bias Temperature Instability Stress in Si p-FinFETs' (IEEE Electron Device Letters, DOI: 10.1109/LED.2020.2992263)

Academic fraud report · Geng Detector

Summary

This AI-assisted review of the 2020 IEEE Electron Device Letters paper (DOI: 10.1109/LED.2020.2992263) by Zhou et al. (Chinese Academy of Sciences, Institute of Microelectronics) concludes a verdict of 'highly suspicious' under the review framework, though all identified anomalies are weak-tier concerns with benign explanations available. No timeline conflicts, methodological contradictions, or statistical red flags were found. Three visual concerns persist at low severity: (1) near-complete overlap of recovery fraction curves across frequencies in Figure 3(b), potentially in tension with the claimed frequency dependence; (2) proportionally scaled trap distribution shapes in Figure 5(a); (3) unusually clean fit in Figure 2(a), consistent with a 4,800-fin parallel large device. A fourth finding — 29% vs. 36% agreement between two independent methods — was judged a positive consistency signal, not evidence of misconduct. Quantitative values (3.37×10¹¹, 2.80×10¹¹, 2.14×10¹¹ cm⁻²) were verified against the source text. Bayesian synthesis yields a 31.3% posterior probability (95% CI [9.9%, 59%], BF=8.64) and an overall low-risk rating. Confidence is limited: no pixel-level image comparison or raw-data statistical testing was possible. Recommended action is to request raw scatter data from the authors; no formal report to editors is warranted at this stage.

Verdict

Highly suspicious (🟠) under the automated review framework — but with important caveats. All retained anomalies are weak-tier (🟡) visual concerns for which benign explanations exist or are partially plausible. No finding constitutes evidence of research misconduct. The overall posterior probability of fraud is 31.3% (95% CI [9.9%, 59%], Bayes factor = 8.64, calibrated against a 5% prior), and the recommended action level is 'no further reporting needed' — current concerns are insufficient to support a PubPeer post or editorial inquiry.

Key findings

Figure 3(b) (⚠️ weak): Recovery fraction curves at 10 Hz, 10 kHz, and 1 MHz appear visually near-identical, in apparent tension with the stated conclusion that "The recovery fraction reduces with an increase in f". A candidate benign explanation: the absolute difference in recovery fraction may be only a few percentage points (ΔN_OT reduction of ~36% on a larger ΔN_IT base), making overlap plausible on normalized axes; the paper also states "ΔI_DCIV demonstrates an f-independent recovery". No quantitative curve data is provided to confirm or refute a systematic offset; raw data verification is needed.
Figure 5(a) (⚠️ weak): ΔD_T energy-distribution curves at three frequencies share peak positions, FWHM, and overall shape, differing only in amplitude — a pattern consistent with either artificial proportional scaling or the paper's own physical conclusion that frequency changes trap *quantity*, not defect energy levels (f-dependence mainly "between E_v and E_i"). Benign explanation is supported by the text; excessive smoothing cannot be visually excluded.
Figure 2(a) (⚠️ weak): Delay-corrected points, measured data, and fit show very low scatter. The devices are 4,800-fin parallel arrays (explicitly stated in the methods; a machine-verification 'not found' error was traced to the 'ﬁ' ligature and line wrapping), which naturally yields low noise. Fits use the established empirical recovery expression of Ref. [21].
Finding 4 (informational, positive signal): ΔN_OT frequency-dependent reduction measured by DCIV (36%, 10 Hz→1 MHz) vs. DMP (29%) on independent instruments (Keysight B1500A vs. B1530) and different devices (4,800 fins vs. 32 fins) match "reasonably"; the reported 7-point discrepancy is more consistent with real than fabricated data.
No evidence of image reuse, timeline conflicts (submitted 2020-03-23, accepted 2020-04-30, published July 2020), unavailability of cited references, or internal methodological contradictions. GRIM/last-digit testing was not feasible given the limited number of in-text numerical values.

Evidence highlights

Extracted values verified present in the source: 10 Hz: 3.37×10¹¹ cm⁻²; 10 kHz: 2.80×10¹¹ cm⁻²; 1 MHz: 2.14×10¹¹ cm⁻²; σ = 1×10⁻¹⁵ cm².
Largest evidence contributions to the posterior: Finding 1 (Figure 3(b) curve overlap, logLR = 1.08, 50.9%) and Finding 2 (Figure 5(a) proportional shape similarity, logLR = 1.08, 49.1%).
One finding (Finding 4) was excluded from synthesis as an informational/consistency item.
The 4,800-fin device claim was manually verified in the source text after an automated check false-negative caused by a typographic 'ﬁ' ligature.

Notes

The 'highly suspicious' label reflects a rule that triggers when ≥3 mutually independent anomalies are retained, not the strength of individual evidence; the authors themselves characterize all retained items as weak and possibly benign.
Bayesian synthesis treats evidence as conditionally independent, which may overstate joint strength for related visual observations.
Recommended follow-up: contact authors for raw scatter data (with error bars) for Figures 3(b) and 5(a); verify whether the three recovery-fraction curves show a resolvable systematic offset on the y-axis.
This report is AI-assisted and for academic discussion only; final determinations of misconduct require institutional investigation, and false positives/negatives are possible.