Geng Report: Concerns Regarding 'Cascaded Improved Neural Network for the Reconstruction, Classification, and Unmixing of the Raman Spectra of Mixed Microplastics' (Anal. Chem. 2026, 98, 8010−8022)

Academic fraud report · Geng Detector

Summary

Verdict: Questionable (yellow). The report flags two main issues in the paper by Huang et al. (DOI: 10.1021/acs.analchem.5c04049). First, in Table 4 (Page 8017), Dataset 9 acquired under the harshest conditions (50 mW laser power, 1000 ms integration time, 50 mJ energy) yields a raw classification accuracy of only 8.97%, with classical preprocessing (SG+AirPLS or WTD+AirPLS) reaching only 15.40% and 17.03%. After the proposed CSAM-ResUNet reconstruction, accuracy reportedly jumps to 82.51%, which the reviewer considers physically implausible and indicative of possible test-set leakage, memorization, or synthetic data fabrication. Second, the authors state that the Raman instrumentation is identical to their prior 2025 Talanta paper (ref 43), with overlapping authorship, suggesting minimal experimental novelty beyond swapping a CNN for a ResUNet variant. No definitive misconduct is proven; the assessment remains a suspicion pending raw-data and code verification.

Verdict

Questionable (🟡) — Two substantive concerns are documented but not independently verified. The central claim of neural reconstruction rescuing classification accuracy from 8.97% to 82.51% under low-SNR conditions is the primary red flag.

Key findings

Implausible accuracy recovery on Dataset 9 (Table 4, Page 8017): 8.97% raw → 15.40% (SG+AirPLS) / 17.03% (WTD+AirPLS) → 82.51% after CSAM-ResUNet.
Possible explanations cited: (1) train/test data leakage, (2) memorization of noise signatures with hallucinated spectral features, (3) reverse-generated/synthetic spectra.
Hardware and methodological reuse: the Raman instrumentation is explicitly acknowledged as identical to the authors' previous work (ref 43, Talanta 2025), with fully overlapping first and corresponding authors.
Incremental model novelty: a CNN in the prior Talanta paper appears to be swapped for an attention-augmented ResUNet (CSAM-ResUNet) and applied to additional plastic mixtures.

Evidence highlights

Table 4, Page 8017 — Dataset 9 (50 mW, 1000 ms, 50 mJ): raw 8.97%; SG+AirPLS 15.40%; WTD+AirPLS 17.03%; CSAM-ResUNet 82.51%.
METHODS, Raman Detection (Page 8014) — Explicit statement that instrumentation is consistent with the authors' previous work [43].
Reference [43] — 2025 Talanta paper sharing first and corresponding authorship with the current work.
DOI: 10.1021/acs.analchem.5c04049.

Notes

The accuracy jump alone does not constitute proof of misconduct; alternative technical causes (e.g., preprocessing-induced bias, evaluation protocol errors) should be considered.
Severity of Finding 1 is rated high (🔴); Finding 2 is rated moderate (🟡) and primarily concerns novelty/output inflation rather than data integrity.
Recommended next steps: request raw spectral files (.txt) and training logs; request side-by-side raw-vs-reconstructed spectra for Dataset 9; consider PubPeer inquiry regarding physical plausibility of recovering chemical features from noise.
The source PDF is referenced as cascaded-improved-neural-network-for-the-reconstruction-classification-and-unmixing-of-the-raman-spectra-of-mixed.pdf.