Integrity concern report: Raman-spectra classification paper with suspected data manipulation and drafting defects

Academic fraud report · Geng Detector

Summary

This report flags the Optics Express paper "Improving the classification performance of microplastics by noise reduction and baseline correction of Raman spectra with a neural network-based algorithm" (DOI: 10.1364/OE.597337) as highly suspicious. The central concern is a physical inversion in the reported classical-algorithm metrics. At constant laser power (29.63 mW), a higher integrated energy dataset (D3: 1250 ms, 37.04 mJ) shows a lower SNR (WTD+AirPLS SNR = 24.07) and lower accuracy (75.57%) than a lower-energy dataset (D4: 1000 ms, 29.63 mJ; SNR = 36.10; accuracy = 83.63%). Since longer integration at fixed power should increase SNR and downstream accuracy, this inversion is physically anomalous. The authors' own SE-ResUNet model does not show the same inversion, suggesting targeted inflation of competing baselines. Additional supporting concerns include a malformed parameter description ("[5,]" with a missing upper bound), an incomplete reference (entry 20 missing authors and title), and apparent overlap of datasets and authors with the group's prior publications, raising data-recycling concerns.

Verdict

🟠 Highly suspicious. The reported classical-algorithm results violate a basic expectation of Raman spectroscopy under fixed laser power: longer integration time (higher deposited energy) should yield higher SNR and accuracy. The authors' own model does not display this inversion, raising the possibility that baseline numbers were tuned to advantage the proposed method. Compounded by drafting defects in formulas and references, the manuscript as a whole warrants editorial and institutional scrutiny.

Key findings

Physical inversion of classical-algorithm metrics (Table 3 & Table 4). For the conventional WTD+AirPLS pipeline, D3 (29.63 mW, 1250 ms, 37.04 mJ) reports SNR = 24.07 and accuracy = 75.57%, whereas D4 (29.63 mW, 1000 ms, 29.63 mJ) reports SNR = 36.10 and accuracy = 83.63%. Lower energy yielding higher SNR and higher accuracy is inconsistent with basic detector physics.
Selective impact. The proposed SE-ResUNet does not display this inversion across D3 vs. D4, consistent with possible selective adjustment of competing-baseline numbers.
Incomplete parameter description (Section 2.3, eq. 4). The explanation of N reads "randomly selected within the range of [5,]," with the upper bound missing.
Broken reference (Ref. 20). Entry 20 begins with a lowercase fragment "dual-tree complex wavelet transform," and is missing author list and article title.
Potential data reuse across the group's publications. References 26 and 27 share first authors (Huang, Chen) with the present study and concern the same microplastic Raman-classification topic, with overlapping size ranges and polymer classes, raising serial-publication concerns.

Evidence highlights

Fixed laser power (29.63 mW); only integration time varies between D3 (1250 ms) and D4 (1000 ms).
Classical-algorithm SNR: D3 = 24.07 < D4 = 36.10 (Table 3).
Classical-algorithm accuracy: D3 = 75.57% < D4 = 83.63% (Table 4).
Drafting defects: parameter range "[5,]" with missing upper bound; Reference 20 lacks authors/title and starts with a lowercase phrase.
DOI: 10.1364/OE.597337.

Notes

The anomalies in Tables 3–4 are presented as reported; the report does not yet contain the raw spectra or processing logs needed to confirm intentional manipulation versus a transcription error. Authors should be asked to release raw spectra and processing scripts for D3 and D4.
The drafting defects (missing upper bound, broken reference) are strong indicators of insufficient proofreading or heavy LLM-assisted drafting, but on their own are not evidence of scientific misconduct.
Overlap of dataset specifications across the group's papers is consistent with legitimate serial studies, but combined with the metric inversion it warrants checking for shared underlying measurements.
Final determination of misconduct requires an institutional or editorial investigation; this report is an alert, not a verdict.