Geng Academic Fraud Detection Report: Modeling the Parasitic Capacitance of Nanosheet FETs With Backside Power Delivery Network (DOI: 10.1109/10.1109/TED.2025.3585478)

Academic fraud report · Geng Detector

Summary

This report assesses "Modeling the Parasitic Capacitance of Nanosheet FETs With Backside Power Delivery Network" (Liu et al., IEEE Transactions on Electron Devices, Vol. 72, No. 9, 2025; DOI: 10.1109/TED.2025.3585478). The overall verdict is YELLOW — suspicious, but leaning innocent; the authors should be asked to verify original simulation data. Copy-move forensics flagged 13 images (img-000 to img-016), but all positives align with fixed grid offsets from multi-panel TCAD/Raphael simulation figures and are assessed as likely false positives. PRNU analysis found two highly homologous image pairs (img-006/012 and img-007/013; NCC=1.000, PCE=128.0) with identical noise statistics (variance mean 62.13/CV=0.850 and 298.6/CV=0.906), most plausibly duplicate extraction of the same figure from the PDF rather than fabrication. The near-perfect model-to-TCAD fit (MAPE ≤ 2.3%) is justified since comparisons involve deterministic numerical simulation rather than noisy experimental measurement. A text AI-generation score of 0.35/1.0 is deemed unreliable and not counted. No evidence of image reuse, data fabrication, statistical anomalies, or timeline inconsistency was found. Confidence is moderate: findings 2 and 3 remain weak-tier unresolved points pending original simulation outputs and PDF inspection.

Verdict

🟡 Suspicious (leaning innocent) — no systematic evidence of fraud; recommend contacting the authors for original simulation data and PDF figure verification.

Paper: Modeling the Parasitic Capacitance of Nanosheet FETs With Backside Power Delivery Network Authors: Hengyi Liu, Sihao Chen, Baokang Peng, Heng Wu, Runsheng Wang, Lining Zhang Journal: IEEE Transactions on Electron Devices, Vol. 72, No. 9, September 2025 DOI: 10.1109/TED.2025.3585478

Key findings

Finding 1 — Copy-move positives across 13 images (img-000 to img-016): Fridrich block matching detected same-offset matched block pairs (mostly at offset (0,32)). All figures are TCAD/Raphael simulation plots, schematics, and multi-panel composites; captions confirm Fig. 6 contains panels (a)–(d), Fig. 7 three panels, Fig. 11 multiple subplots. Regular repeated elements (axes, borders, tick marks) naturally produce same-offset matches at fixed grid offsets. Assessed as likely false positive.
Finding 2 — PRNU highly homologous pairs (img-006/012, img-007/013): NCC=1.000, PCE=128.0. Noise variance data ([I10]/[I20], [I12]/[I22]) are identical item-by-item (variance mean 62.13/CV=0.850; variance mean 298.6/CV=0.906). Most plausible explanation: the same figure was extracted/rendered twice from the PDF, not two "independent experiments" — though this benign explanation is speculative and not verbatim-supported in the text. Retained as weak-tier concern; verify against original PDF.
Finding 3 — Suspiciously perfect model–simulation fit: Model vs TCAD points in Fig. 6–8 and the 5-RO transient waveforms in Fig. 11(f) show near point-by-point overlap. The paper states "the mean absolute percentage error (MAPE) of the model is not greater than 2.3%" (verified in text). Comparison is against deterministic TCAD simulation (Raphael Poisson solver, per the methods section), not noisy experimental measurement, giving the high agreement a reasonable physical basis. However, pixel-level agreement including minor noise/phase jitter in thousands-of-device mixed-mode circuit simulation remains technically worth questioning. Retained as weak-tier concern.
Finding 4 — AI-generated text score 0.35/1.0: Below usable threshold; templated engineering-writing style causes high false-positive rates on short texts. Not counted as an anomaly.

Evidence highlights

Copy-move offsets (e.g., (0,32), with 33/59/34 matched block pairs) align with multi-panel layout grids.
PRNU: NCC=1.000, PCE=128.0 for both pairs; noise statistics match exactly (62.13/CV=0.850 and 298.6/CV=0.906).
MAPE ≤ 2.3% claim verified against the paper text.
Not triggered: No image reuse across experiments (pattern 1); no numeric tables available for GRIM/last-digit testing (key figures: 19.4%, 14.1%, 7.4%, MAPE 2.3% — too few values; no SD/p-values reported); no hypothesis tests; timeline self-consistent (submitted 2025-05-28, accepted 2025-06-26, online 2025-07-17; all citations predate submission; review date 2026-10-12); tools/resources (Raphael, BSIM-CMG, IRDS) are real.
Programmatic verification notes: "not found in text" flags for machine-forensics outputs ([I2]–[I27], [T1]) are verification-scope omissions, as these values come from forensic tool outputs, not the paper body.

Notes

Limitations: Findings 2 and 3 have plausible benign explanations but lack positive verbatim support in the paper; both are downgraded to weak-tier and retained. Final determination requires the original simulation output files and inspection of the source PDF for duplicated page figures.
Recommended actions:
Contact authors for raw TCAD/Raphael simulation outputs and 5-RO mixed-mode waveform data.
Verify whether img-006/012 and img-007/013 are duplicate renderings of the same page/figure in the source PDF.
No journal/institutional report needed at this time (no systematic fraud evidence found).
Disclaimer: AI-assisted report for academic discussion only. Final determination of misconduct requires professional institutional investigation; false positives and false negatives are possible. Confidence: moderate; verdict leaning innocent pending author response.