Integrity Review Report: Electromigration of Backside Power Delivery Networks for PPA-Reliability Tradeoffs at N2 Node (IEDM 2023, DOI: 10.1109/IEDM45741.2023.10413884)

Academic fraud report · Geng Detector

Summary

This report assesses the IEDM 2023 conference paper 'Electromigration of Backside Power Delivery Networks for PPA-Reliability Tradeoffs at N2 Node' (DOI: 10.1109/IEDM45741.2023.10413884), a purely simulation-based study (coupled electro-thermal-stress KMC method). Overall verdict: highly suspicious, with anomalies concentrated in figure rendering and statistical presentation rather than classic fabrication hotspots. Finding 1 (orange): Fig. 7 and Fig. 8 microstructure simulation images show nearly identical grain-boundary textures despite captions describing four different feature structures, suggesting reuse of the same 3D model base. Findings 2-3 (yellow): Fig. 16 error bars are suspiciously uniform across materials, and Fig. 14(b) Lognormal plots are 'textbook perfect', though benign explanations (same KMC framework, simulation variance) partially apply. Finding 4 (yellow): Table II shows constant column differences flagged by automated statistics, but the PDF's corrupted character encoding prevented verbatim verification. The report's automated Bayesian synthesis states a 99.2% posterior, but this rests partly on weak text-metric and noise-variance evidence; independent forensic confirmation and original KMC output from the authors are required before any misconduct determination.

Verdict

Highly suspicious (orange/yellow composite). The paper is purely computational (coupled electro-thermal-stress Kinetic Monte Carlo simulation of backside power delivery network electromigration), so classic fabrication hotspots (blots, gels, microscopy) do not apply. The most serious anomaly is apparent reuse of an identical microstructure render across four nominally different feature structures (Fig. 7 vs. Fig. 8). Additional statistical-presentation anomalies (Figs. 14b, 16; Table II) are suggestive but carry plausible benign explanations. No definitive misconduct finding; original simulation output is required to resolve.

Note: the PDF's text layer has a systematic character-encoding offset, which caused all programmatic verbatim checks to fail. After decoding, key data points cited by the report ("100 samples", "ΔR 10%", "Lognormal", "6.22", "3.38", "8%", the four feature structures in Fig. 7/8 captions) were confirmed present in the original text — a machine-verification miss, not data fabrication.

Key findings

Finding 1 (Fig. 7 / Fig. 8, severity: orange — upheld): Gray-grid microstructure backgrounds (noise pattern, grain boundary orientation, shading) are nearly identical across Fig. 7 (BPR-1: W M0A / W VBPR; BPR-2: W nTSV / Cu BSM) and Fig. 8 (TSVM-1: W M0A / W TSVM; TSVM-2: W TSVM / Cu BSM) — four different structures/material systems. Stochastic KMC grain structures should not be texture-identical across W/Cu and nTSV systems. Consistent with forensic copy-move hits (e.g., offset (40,0) with 13 pairs; (32,0) with 13/17 matched block pairs).
Finding 2 (Fig. 16, severity: yellow — benign explanation partially applies): TTF comparison error bars are nearly identical across Mo, Ru, W groups, despite lognormal TTF distributions from 100 independent KMC runs (paper states "TTF of BSC is 6.22× longer than BPR", "Mo metal is 3.38× longer than W", "error smaller than 8%"). Downgraded one level because a single KMC framework may naturally produce regular variance.
Finding 3 (Fig. 14(b), severity: yellow — benign explanation possible): Lognormal probability plots for W/Ru/Mo show near-perfect parallel fitted lines with 100 samples each; such plots usually show end deviations and local zig-zag. Alone insufficient for any conclusion.
Finding 4 (Table II, severity: yellow — insufficient basis): Automated statistic (S1) flagged constant inter-column differences. Table headers/entries confirmed via decoding, but numeric values could not be reliably transcribed for verification due to encoding corruption. Recorded as a lead only, not as evidence of misconduct.

Evidence highlights

Fig. 7/8 captions (verified post-decoding) claim four distinct feature structures; the rendered textures are visually near-identical, with no physical explanation offered.
Forensic copy-move matches: offset (40,0) 13 pairs; offset (32,0) 13/17 matched block pairs — directionally consistent with visual observation.
Verified textual data points: "100 samples", ΔR 10% TTF criterion, Lognormal distribution, 6.22× (BSC vs. BPR), 3.38× (Mo vs. W), <8% error.
Automated Bayesian synthesis (auto-generated, not a misconduct finding): posterior 99.2% (95% CI [93%, 99.9%]), BF = 2.49e+03; top contributors were the weak AI-text metric (T1, logLR=1.97, 26.4%), the unverified Table II column statistic (S1, logLR=1.87, 23.9%), and noise-variance image flags (I36/I31/I11, logLR=0.99 each, ~13%/12.7%/12.6%). The synthesis assumes conditional independence and likely overstates correlated evidence; treat with caution.
Timeline check: no temporal inconsistencies (©2023 IEEE, references 2011–2023). Internal methodological consistency between abstract, methods, and conclusions confirmed.

Notes

Uncertainty: Findings 2–4 retain plausible benign explanations; only Finding 1 is judged "upheld" and even it concerns render reuse rather than fabricated results per se.
The 99.2% automated posterior should not be read as a misconduct probability; its top contributors include weak signals (mild AI-text feature 0.15/1.0, judged informational).
Recommended follow-up: request original 3D/KMC output files for Figs. 7–8 and raw statistical scripts/data for Figs. 14/16; re-verify Table II values from an intact PDF; consider raising on PubPeer and with IEDM/IEEE editors pending author response.
Limitation: source PDF text layer is systematically character-shifted, degrading automated verification throughout.

*AI-assisted report for academic discussion only; final determination of misconduct requires institutional investigation.*