Geng Academic Integrity Review: Electromigration of Backside Power Delivery Networks for PPA-Reliability Tradeoffs at N2 Node (IEDM 2023)

Academic fraud report · Geng Detector

Summary

This AI-assisted review of the IEDM 2023 paper 'Electromigration of Backside Power Delivery Networks for PPA-Reliability Tradeoffs at N2 Node' (DOI: 10.1109/IEDM45741.2023.10413884, Cai et al., Sun Yat-sen University) concludes a verdict of 'questionable' (yellow flag). The paper is purely computational (coupled electro-thermal-stress KMC/electromigration simulation) with no wet-lab imagery. The original automated evidence pass produced many false 'not found' results due to PDF font-encoding corruption; on recheck, key data points (BPR-1, TSVM-1, 3.28X, 6.22X, 100 samples, Table II values 0.51, 0.59, 0.61, 1.38, 0.90) were verified in the source. Four findings were assessed: (1) high visual similarity between Figs. 7 and 8 3D renders despite depicting voids at physically different locations — insufficient evidence, moderate confidence only, awaiting original simulation snapshots; (2) overly neat lifetime multipliers (3.28X, 6.22X) — benign, expected for 100-sample simulation statistics; (3) mass copy-move/noise forensic alerts — cleared, explainable by chart-type figures and IEEE PDF compression; (4) constant column differences in Table II — cleared, cited physical constants from refs [14][15] violate test assumptions. No fabrication evidence found; the Fig. 7/8 similarity remains the sole open question.

Verdict

🟡 Questionable (存疑) — no evidence of data fabrication; one unresolved visual-similarity question (Fig. 7 vs. Fig. 8) pending original simulation data.

The paper is a pure-simulation study (fully coupled electro-thermal-stress KMC/EM simulation) with no wet-lab images (Western blots, gels, micrographs). All figures are schematics, 3D simulation renders, or statistical charts, as confirmed by captions in the source text. Note: the original PDF suffered font-encoding character shifts, causing machine-verification false negatives; the recheck restored key data points (BPR-1, TSVM-1, VBPR, TSVM, power line GND, 100 Samples, 3.28X, 6.22X, 20.43, 32.10, and Table II values 0.51, 0.59, 0.61, 1.38, 0.90), confirming findings 1, 2, and 4.

Key findings

Finding 1 — Fig. 7 vs. Fig. 8 similarity (🟡, insufficient evidence): The 3D renders (BPR-1 vs. TSVM-1 void evolution, both connected to power line GND) reportedly share near-identical geometry, lighting, shadows, and bottom texture at moderate visual confidence. The captions confirm different structures, and the text places voids at different physical locations (VBPR bottom for BPR-1; MoA/TSVM interface for TSVM-1), creating tension with the visual similarity. Benign explanation (unified render pipeline on the same simulation platform) is plausible but unverified; authors should provide original simulation snapshots.
Finding 2 — overly neat lifetime multipliers in Fig. 16 (🟡, benign explanation supported): BPR(Mo) = 3.28X of BPR(W); BSC(W) = 6.22X of BPR(W). Values verified in source. The paper explicitly reports TTF distributions from 100-sample simulation statistics (Fig. 10); precise ratios are normal for simulated means. Only note: confidence intervals/dispersion were not reported.
Finding 3 — mass copy-move / noise-variance alerts (🟡, cleared): 20+ of 40 embedded images flagged (e.g., I2, I5, I6, I9, I10, I15, I20, I26, I32, I34–I37, I39, I42, I45, I46) with same-offset block matches such as (32,0), (0,32), (40,0). All captions confirm schematic/render/chart figure types; regular repeated elements (axes, borders, legends) and IEEE two-column PDF compression systematically explain the alerts. No high-risk image types present; no alert indicates content-level manipulation.
Finding 4 — constant column differences (S1) in Table II (🟡, cleared): Table II contains physical constants (Ea, Eb, E_m, k, resistivity) cited from refs [14][15]; the column_diff test assumes experimental measurements and is inapplicable to cited literature constants.

Evidence highlights

DOI: 10.1109/IEDM45741.2023.10413884 (IEDM 2023, IEEE, published December 2023)
Verified data points: 3.28X, 6.22X, 100 Samples, 20.43, 32.10; Table II values 0.51, 0.59, 0.61, 1.38, 0.90
Forensic triage: S1 (column_diff) cleared — test assumptions not met; T1 (AI-text score 0.15/1.0) cleared — negligible score, short conference paper; I1–I48 (copy-move/noise) cleared — chart-type figures and PDF compression pipeline provide source-text-supported benign explanations
Not triggered: blot reuse/splicing (no biological imagery); p-hacking (no p-values reported); timeline (no reagent/market timelines); methodology (KMC parameters internally consistent with refs [11]–[16])

Notes

Limitations: the Fig. 7/8 similarity assessment rests on a moderate-confidence visual summary, not pixel-level forensic confirmation; a benign rendering-pipeline explanation cannot be confirmed or excluded from the text alone. PDF encoding corruption required manual restoration of machine-verified data points.
Recommended actions: request original simulation output files and KMC parameters for Figs. 7/8 from the authors; optionally raise a neutral question on PubPeer; no journal/institutional report warranted at this stage (below threshold).
This report is AI-assisted, for academic discussion only; final misconduct determinations require institutional investigation.