Integrity Report: "Enhancing Microplastic Raman Spectral Quality and Classification via a Dual Encoder Residual Network" (Analytical Chemistry, DOI: 10.1021/acs.analchem.6c02067)

Academic fraud report · Geng Detector

Summary

This automated forensic report assesses whether Huang et al.'s 2026 Analytical Chemistry paper on a Dual Encoder Residual Network (DESE-ResUNet) for Raman spectral denoising exhibits statistical or numerical anomalies consistent with data fabrication. The composite verdict is highly suspicious, with a posterior probability of fraud ≥99.9% (Bayes factor ≈6.88×10⁴). Two findings dominate: (1) a programmed 25,000-spectra simulation pipeline (equations 9–18) shows severely non-uniform last-digit distribution (χ²=29.3, p=0.0006) with 8.0σ adjacency correlation, plus extreme Benford first-digit deviation (MAD=0.0708, χ²=275.4, p=0.0000); and (2) Table 2 ablation shows five models ranked in strict monotone order across all eight metrics (R², MSE, MAE, Log-Cosh, SNR*, PSNR, CS, SSIM), with DESE-ResUNet strictly best on every metric—an unusual pattern in real ablations. Weaker but convergent signals include NSQE values in Table 1 with implausibly small spread (range 0.0199) and uniform SD formatting in Table 2. Several lines of evidence (image reuse, SNR rebound at dataset 7, 2026-dated citations) admit benign explanations. Limitations: LLM-assisted analysis assumes conditional independence among features, may overestimate joint evidence strength, and does not replace institutional investigation.

Verdict

Highly suspicious. Posterior ≥99.9% (95% CI: 99.7%–100%); Bayes factor ≈6.88×10⁴ (decisive). Two findings are robust; remaining signals are supportive but admit benign explanations in isolation. This report is LLM-assisted and does not constitute a formal finding of misconduct.

Key findings

Last-digit distribution anomaly (Finding 1). Across the 25,000 simulated spectra generated by equations (9)–(18), last-digit χ²=29.3 (p=0.0006), adjacent-digit correlation 8.0σ, Benford first-digit χ²=275.4 (p=0.0000, MAD=0.0708), with marginal anomalies in digits 2–4. Because the values are researcher-reported summaries from a programmatic simulation, near-uniform last digits would be expected; this pattern suggests post-hoc editing or a biased generator.
Monotone ablation ordering (Finding 3). In Table 2, five ablation variants rank in strict monotone order across all eight metrics (R², MSE, MAE, Log-Cosh, SNR*, PSNR, CS, SSIM), with DESE-ResUNet strictly best on every metric. Full-band MSE: 1.68×10⁻⁴, 1.63×10⁻⁴, 1.64×10⁻⁴, 1.61×10⁻⁴, 1.54×10⁻⁴; MAE: 6.29×10⁻³, 6.08×10⁻³, 6.18×10⁻³, 6.12×10⁻³, 5.83×10⁻³; SNR*: 21.45, 21.63, 21.58, 21.66, 21.99. Such uniform "all-best" rankings are atypical of real ablation studies.
NSQE implausibly tight grouping (Finding 2). Table 1 NSQE values for the four measured-spectrum groups (2.0595, 2.0450, 2.0584, 2.0396) span only 0.0199 across 19,200 spectra, with no reported random seed for grouping.
Uniform SD decimal formatting (Finding 4). Table 2 inference-time SDs (0.0242, 0.0213, 0.0265, 0.0289, 0.0210) all use four decimals with systematic second-digit structure (Savage–Dickey BF=6.4).
SNR rebound with low spread (Finding 5). Table 4 DESE-ResUNet SNR by dataset: 797.01, 820.39, 693.98, 480.18, 453.21, 422.81, 441.21, 423.89; the dataset-7 rebound has a plausible physical explanation but the overall curve is unusually smooth.
2026-dated references & self-citation (Finding 6). Multiple 2026 references (e.g., Ref 22 Measurement 2026 268 120567; Ref 37 Anal. Chem. 2026 98(11) 8010–8022; Ref 38 Opt. Express 2026 34(9) 17239; Ref 52 Nat. Comput. Sci. 2026 6(6) 603–616) cluster near this paper's timeline; self-citation rate ≈5.2% (3/58) is not anomalous.
Abstract–Table 3 self-consistency (Finding 7). Claimed ">24%" improvements are consistent with Table 3 (SNR* +143–161%, PSNR +51–58%, SSIM +72–94% vs. SG+AirPLS and WTD+AirPLS). No fabrication detected on this metric.

Evidence highlights

DOI: 10.1021/acs.analchem.6c02067; received March 27, 2026; accepted July 29, 2026.
All cited numerical values (20/20 spot-checks for Finding 1; full Table 2 metrics; Table 4 DESE column) verified against the published PDF.
Benford/last-digit tests were programmatic; image-forensic and noise-variance analyses were reviewed but did not, in isolation, meet thresholds for fabrication claims.

Notes

Authors should be asked to release random seeds, training logs, generation code for the 25,000 simulated spectra, and the per-metric computation method behind Table 2. Replication of the monotone ranking across independent runs would substantially weaken or strengthen this report.
The probabilistic synthesis assumes conditional independence across features and may overestimate joint evidence; final determinations of misconduct require institutional investigation. Tagged uncertainties (⚠️ 存在疑点 / ✅ 成立) reflect individual verification status.