Geng Integrity Assessment: 'Cascaded Improved Neural Network for the Reconstruction, Classification, and Unmixing of the Raman Spectra of Mixed Microplastics' (Anal. Chem., DOI: 10.1021/acs.analchem.5c04049)

Academic fraud report · Geng Detector

Summary

Verdict: Questionable (🟡). The paper, a machine-learning/Raman spectroscopy study by Huang et al., shows no systemic fraud at the text, table, or equipment-timeline level. Internal arithmetic in Tables 3 and 4 is consistent, and the reported GPU (NVIDIA RTX 4070 Ti Super, released January 2024) is anachronistically compatible with the July 2025 submission. The authors publicly released code on GitHub and used several public datasets, which is a positive integrity signal. The principal concern is that under extremely degraded conditions (Dataset 9: 50 mW laser, 1 s integration, ~50 mJ received energy) raw-spectra classification collapses to 8.97% (near random), yet post-reconstruction classification reaches 82.51%—an implausibly large gain suggestive of overfitting or possible train/test data leakage. Pixel-level image forensics could not be performed because only text and tabular content were supplied. Confidence is moderate; findings hinge on reproducing the authors' open-source pipeline.

Verdict

🟡 Questionable. No definitive fraud detected, but one high-suspicion performance claim warrants independent code-based verification.

Key findings

Suspiciously strong reconstruction under extreme noise: In Dataset 9 (50 mW laser power, 1 s integration, ~50 mJ energy), raw-spectra classification accuracy collapses to 8.97% (near random for the multi-class problem), while the cascaded network's reconstructed-spectra accuracy reaches 82.51%. The magnitude of this recovery is implausible and may indicate overfitting or train/test data leakage (Data Leakage).
Internal arithmetic is self-consistent: Tables 3 and 4 (Page 8017) compute correctly across columns and rows; no fabricated-random-number pattern detected.
No anachronistic equipment: The reported NVIDIA GeForce RTX 4070 Ti Super GPU was released in January 2024, well before the July 3, 2025 submission date—timeline logic is sound.
Open-code and open-data practice: Code is released at https://github.com/V1S10NAL/CSAM-ResUNet, and the work is validated on multiple public Raman datasets (References 44–48). Authors also apply combinatorial logic C₇² = 21 for mixed-microplastic categories, indicating methodological rigor.
Pixel-level image forensics not executable: Source figures (Figures 1–8) and raw image pixels were not provided, so Geng's First Form (image reuse detection) and Third Form (image splicing detection) could not be applied.

Evidence highlights

Table 3 / Table 4 (Page 8017): Dataset 9 baseline = 8.97%; reconstructed = 82.51%.
Page 8013 (Model Training): explicit mention of RTX 4070 Ti Super.
Page 8012 / Page 8016 (Neural Network Structure): GitHub link https://github.com/V1S10NAL/CSAM-ResUNet.
Combinatorial derivation: C₇² = 21 mixture classes (Pages covering dataset construction).
DOI: 10.1021/acs.analchem.5c04049.

Notes

The suspicious Dataset 9 result is the sole substantive concern; the authors do acknowledge the difficulty but do not rule out leakage.
Recommended next actions: (1) clone the GitHub repository and verify test-set partitioning for Dataset 9; (2) retrain under strict held-out splits; (3) optionally raise the over-idealized denoising-classification behavior on PubPeer for community discussion.
Limitations of this assessment: no pixel-level image analysis was possible; numerical claims rely on the as-published Tables 3–4 only.
Disclaimer (preserved from original): this report is AI-assisted and intended for academic discussion only; final misconduct determinations require official institutional investigation.