Fraud Analysis Report: Joint Optimization of Layout and Electrolyzer Capacity for Decentralized Offshore Wind-Powered Hydrogen Production (DOI: 10.1109/TII.2025.3538067)

Academic fraud report · Geng Detector

Summary

This report evaluates the IEEE Transactions on Industrial Informatics paper by Du, Shen, Lu, and Ding (2025), which proposes a BLDDM algorithm for offshore wind-powered hydrogen production. Overall verdict: highly suspicious. The primary concern is the implausible numerical equivalence between baseline algorithms. In Table III, PSO and NMSM report identical mean runtimes of exactly 1852.3s, identical LCOH values of 7.56 €/kg, and identical standard deviations of 0.00, which is statistically improbable given their fundamentally different search mechanisms. Table V (72 platforms) reproduces nearly identical runtimes (15643.2s vs 15642.8s) and identical LCOH values (9.12 €/kg). A secondary concern is a textual inconsistency: the paper claims Case 4 reduces LCOH by 15.82% versus Case 2, but Table II yields only ~12.07%; the 3.21% claim versus Case 3 checks out. Timeline and image reuse checks raised no flags, though pixel-level image analysis was not possible. Confidence is moderate-to-high for the data anomaly but lower absent the authors' raw logs.

Verdict

Highly suspicious. Two independent, serious anomalies were identified: (1) implausible numerical equivalence across structurally different optimization algorithms in Tables III and V, and (2) a numerical inconsistency between narrative claims and tabulated values in Section V.B. The combination of these issues warrants formal author response and editorial investigation.

Key findings

Impossible baseline equivalence (Tables III & V): PSO (global stochastic) and NMSM (local simplex with warm-started initial point) report identical mean runtimes (1852.3 s), identical LCOH means (7.56 €/kg), and identical standard deviations (0.00) in Table III. In Table V (72 platforms), runtimes differ by only 0.4 s (15643.2 vs 15642.8) and LCOH values are identical at 9.12 €/kg. Such convergence across different algorithmic paradigms on a non-convex problem has probability approaching zero.
Text–table inconsistency (Section V.B, p. 4165): The paper states Case 4 reduces LCOH by 15.82% versus Case 2. Using Table II values (Case 2 = 7.54 €/kg, Case 4 = 6.63 €/kg), the actual reduction is (7.54 − 6.63)/7.54 ≈ 12.07%, not 15.82%. The 3.21% claim versus Case 3 (6.85 → 6.63) does verify.
Timeline plausibility: No temporal anomalies. Submission 2024-10-22; software Matlab R2019a; Intel i5-12500 (2022); download header September 02, 2025 — internally consistent.
Image reuse: Pixel-level analysis was not performed because only text-extracted content was available. No textual contradictions between captions (Figs. 4–8) and the surrounding prose were found.

Evidence highlights

Table III, p. 4165: PSO vs NMSM runtime = 1852.3 s vs 1852.3 s; LCOH = 7.56 €/kg vs 7.56 €/kg; SD = 0.00 vs 0.00; SQP likewise identical.
Table V, p. 4166: PSO vs NMSM runtime = 15643.2 s vs 15642.8 s; LCOH = 9.12 €/kg vs 9.12 €/kg.
Table II values used for cross-check: Case 2 LCOH = 7.54 €/kg, Case 3 = 6.85 €/kg, Case 4 = 6.63 €/kg.
Manuscript-stated reductions: 15.82% (vs Case 2) — does not match recomputed 12.07%; 3.21% (vs Case 3) — matches.
DOI: 10.1109/TII.2025.3538067.

Notes

These findings indicate probable data fabrication or copy-paste assembly of baseline comparisons, but they do not constitute proof of misconduct on their own. Definitive determination requires the authors' raw run logs and convergence traces for PSO, NMSM, and SQP across all reported cases.
The 15.82% vs 12.07% mismatch could alternatively reflect a drafting oversight where the table was updated but the prose was not, though this still represents a material error in a published, peer-reviewed article.
Image-based checks were constrained by input format; high-resolution figures should be re-examined.
All percentages, runtimes, and DOI above are reproduced exactly as reported by the source report or paper.