Integrity Review: 'Improving the classification performance of microplastics by noise reduction and baseline correction of Raman spectra with a neural network-based algorithm' (Optics Express, DOI: 10.1364/OE.597337)

Academic fraud report · Geng Detector

Summary

Verdict: Highly suspicious. The review identifies multiple serious internal inconsistencies and physically implausible results in the paper by Jiajin Chen et al. published in Optics Express (Vol. 34, No. 9, 2026). First, a clear arithmetic contradiction is found in Section 3.4: the authors state the dataset comprises 1,800 Raman spectra partitioned in a 7:2:1 ratio, yet subsequently describe a test set of 300 spectra. Under 7:2:1, 10% of 1,800 yields 180 test samples, not 300; a 300-sample test set would imply a total of 3,000 spectra. Second, Tables 3 and 4 contain data that violates physical expectations: as integration time/laser energy decreases from Dataset 4 (1000 ms) to Dataset 5 (750 ms), input signal quality deteriorates, yet the SE-ResUNet model's reported SNR rises anomalously from 1446.07 to 1857.15 and accuracy rises from 95.10% to 96.90%, suggesting data fabrication or table-entry errors. Third, the authors show a pattern of rapid, structurally similar publications on microplastics, Raman spectroscopy, and neural networks. Image-based checks could not be performed. Confidence in findings 1 and 2 is high; findings require official investigation.

Verdict

Highly suspicious. The paper contains an unambiguous arithmetic inconsistency and physically implausible quantitative results that are consistent with data fabrication or careless table construction. Multiple independent concerns raise serious doubts about the reliability of the reported experiments.

Key findings

Mathematical contradiction in dataset partitioning (Section 3.4, Page 17249): The authors state a total of 1,800 Raman spectra partitioned in a 7:2:1 ratio, then later describe a test set of 300 spectra. Under 7:2:1, the test set should be 180 spectra (10% of 1,800). A 300-sample test set would require a total dataset of 3,000 spectra. The two statements cannot both be correct.
Physically implausible SNR/accuracy trends (Tables 3 and 4, Pages 17249–17250): Moving from Dataset 4 (1000 ms integration) to Dataset 5 (750 ms), input signal quality worsens. The conventional method (WTD+AirPLS) reflects this, with SNR dropping from 36.10 to 25.42 and accuracy dropping from 72.43% to 55.70%. However, the SE-ResUNet results move in the opposite direction: SNR increases from 1446.07 to 1857.15, and accuracy rises from 95.10% to 96.90%. A denoising model producing higher output SNR from lower-quality input is not explainable without hallucinated features.
Sausage-paper pattern of related outputs (References 26, 27): Multiple closely related papers by the same group on microplastic Raman spectroscopy with neural networks have appeared in 2025–2026, indicating a templated production pattern consistent with rapid, framework-recycled publications.
Image integrity unverified: Figures 1–7 (spectral plots, training curves, heatmaps) could not be examined at the pixel level because only text was available for review.

Evidence highlights

Arithmetic contradiction: "The dataset was partitioned into training, validation, and testing subsets according to a 7:2:1 ratio" versus "classify a test set of 300 spectra"; 10% of 1,800 = 180, not 300.
Reversed SNR trend (Table 3/4): SE-ResUNet SNR: Dataset 4 = 1446.07 → Dataset 5 = 1857.15; SE-ResUNet accuracy: 95.10% → 96.90%, opposite to the conventional-method trend (36.10 → 25.42 SNR; 72.43% → 55.70% accuracy).
Publication pattern: References [26] (2026) and [27] (2025) by the same author group on microplastics + Raman + neural networks.

Notes

DOI: 10.1364/OE.597337
Source PDF: oe-34-9-17239.pdf
Image-based forensic checks (duplication, splicing, axis cropping) were not possible from the available text-only materials and remain pending.
All quantitative figures above are reproduced exactly as reported in the source review; no values have been inferred or modified.
The current assessment is based solely on the textual content and tabulated data provided. Definitive determination of misconduct requires an official institutional investigation, including inspection of raw spectra, training logs, and partitioning code.