Integrity Review 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

Verdict: Questionable (low-to-moderate suspicion). The authors, including Weixiang Huang, Jiajin Chen, Hao Xiong, Ligang Shao, Tu Tan, Guishi Wang, Kun Liu, and Xiaoming Gao, report a dual-encoder residual network for Raman microplastic preprocessing. After programmatic re-verification, most quantitative allegations in the original Chinese detection report (specific χ² values, exact copy-move block counts, named panel labels) could not be located in the published manuscript and were accordingly withdrawn. Surviving signals are weaker: (1) integer-pixel-offset copy-move block matches detected by Fridrich analysis in 8 of 10 embedded figures, where benign explanations (axes, borders, stacked-lane geometry) cannot be ruled out without source files; (2) implausibly high precision in Table 2's inference-time standard deviations (e.g., 0.0242, 0.0213, 0.0265) given n=5 independent runs, although instrument/software auto-reporting may explain this; (3) visual inspection suggesting excessive morphological similarity across panels of Figures 6 and 7 stacked spectra at varying noise/energy conditions, raising questions about whether Raw spectra were independently acquired. Confidence is limited; raw data disclosure is required before any definitive judgment.

Verdict

🟡 Questionable — Multiple weak signals warrant clarification but do not constitute evidence of misconduct. The original report's most damning specific numbers (χ², exact block-match counts, named sub-panel codes) were not found in the source PDF and are therefore not admissible. Remaining concerns are circumstantial and require raw data for resolution.

Key findings

Image forensics (copy-move, weak): Fridrich-style block matching flagged integer-pixel-offset duplicated regions in 8 of 10 embedded images (covering Figures 2–8), including offsets (32,0), (0,32), (0,88), (64,0), (136,0), (0,80), (88,0), (80,0). Such integer offsets commonly arise from regular repeated graphical elements (axes, borders, tick marks, stacked-lane separators) and may be benign.
Statistics (last-digit anomaly, weak): Table 2 reports five model inference times of 0.6987±0.0242, 0.7777±0.0213, 0.9545±0.0265, 0.9995±0.0289, and 0.9776±0.0210 across n=5 independent runs. SDs reported to 4 decimal places for n=5 are statistically unusual but can plausibly reflect instrument or software auto-reported precision.
Figure 6 (PVC, 8 noise levels), weak: Panels (a–h) under varying experimental noise (laser power stepping from ~780 mJ to ~30 mJ) show Raw/SG/AirPLS/WTD/AirPLS/DESE-ResUNet traces that appear highly similar across noise levels. The paper does not clarify whether each panel used independently acquired Raw spectra or rescaled/duplicated noise samples.
Figure 7 (8 polymers, 3 energy conditions), weak: Three measured datasets (datasets 2, 5, 8) with conditions 156.14 mW/3000 ms, 29.63 mW/2500 ms, 29.63 mW/1000 ms (Table 4) are described. Panels show Raw baseline fluctuation patterns that appear highly reproducible across energy levels, again suggesting possible reuse of scaled noise samples. Independent verification requires the original unaveraged spectra.

Evidence highlights

DOI: 10.1021/acs.analchem.6c02067 — Analytical Chemistry (ACS); received March 27, 2026; revised July 28, 2026; accepted July 29, 2026.
Table 2 inference-time entries: 0.6987±0.0242, 0.7777±0.0213, 0.9545±0.0265, 0.9995±0.0289, 0.9776±0.0210 (mean ± SD across n=5 independent runs).
Figure 6 caption (paraphrased): performance comparison between the neural network and traditional algorithms for PVC spectral baseline correction and noise removal under varying experimental noise levels.
Figure 7 caption (paraphrased): effectiveness of the neural network in baseline correction and noise removal for various microplastic Raman spectra under different experimental conditions; references measured datasets 2, 5, 8.
Table 4 laser/energy conditions: 156.14 mW / 3000 ms; 29.63 mW / 2500 ms; 29.63 mW / 1000 ms.
Programmatic statistics cited (χ²=29.3, p=0.0006; last-digit correlation deviation = 8.0σ) are computer-generated and do not appear in the manuscript.
Specific sub-panel codes invoked in the original report (e.g., "WTD_db4_3+AirPLS_10^3", "SG_11_2+AirPLS_10^3", explicit panel-letter lists) were not located in the source PDF.

Notes

All quantitative statistics drawn from copy-move/PRNU/Benford-style analyses carry known applicability constraints (image format, compression, figure composition) and should be interpreted as leads rather than proof.
Benford/last-digit analyses are highly sensitive to sample size, rounding rules, and data origin (measured vs. computed), and n=5 is too small for robust distributional claims.
The authors (Weixiang Huang, Jiajin Chen, Hao Xiong, Ligang Shao, Tu Tan, Guishi Wang, Kun Liu, Xiaoming Gao) have not provided a public response in the materials reviewed.
Recommended next steps: request original unaveraged spectra (.csv/.txt) for Figures 6 and 7; request raw timing logs underlying Table 2; request pre-assembly source files and plotting scripts for Figures 2, 6, and 7; submit a neutral technical query on PubPeer. A formal misconduct allegation is not currently warranted.