Geng integrity report — A Lightweight Multi-Task Learning Framework for Battery Degradation Prediction with Limited Data (DOI: 10.1109/TTE.2025.3591534)

Academic fraud report · Geng Detector

Summary

Verdict: Doubtful (yellow flag). No conclusive evidence of intentional data fabrication or image manipulation was identified; the study relies entirely on well-known public battery datasets (Severson 2019, Attia 2020, Zhu 2022), which removes the typical risk of fabricated experimental data. However, multiple low-level but visible editorial and scholarly-care issues undermine the paper's rigor. Key findings include (1) duplicate/overlapping equation numbering, where equations (7)–(15) used in Section III.C reappear as (7)–(9) for evaluation metrics in Section IV.A, indicating copy-paste from another source; and (2) a misattributed citation for the Adam optimizer, which is cited as reference [20] rather than the canonical Kingma & Ba (2014, arXiv:1412.6980). Other checks—software timeline (MATLAB R2024a), dataset provenance, and statistical plausibility of Table IV values—were unremarkable. Confidence is moderate; the issues strongly suggest careless writing or partial copy-paste rather than orchestrated fraud, but they justify a PubPeer comment and an Erratum.

Verdict

Doubtful (🟡). No firm evidence of research misconduct was detected. Several editorial and citation errors are serious enough to warrant a public comment and author correction, but they are consistent with rushed/poorly proofread writing rather than deliberate fraud.

Key findings

Equation numbering conflict (Equation reuse). Section III.C uses equations (7)–(15) for projection/compression derivations, while Section IV.A restarts numbering at (7) for MAPE, (8) for RRMSE, and (9) for R².
Misattributed optimizer citation. The Adam optimizer is cited as reference [20], a 2025 IEEE TTE paper on battery lifetime prediction. The canonical source (Kingma & Ba, 2014; arXiv:1412.6980) is missing.
Public dataset usage. Four datasets (124 LFP, 45 LFP, 28 NMC, 28 NCA cells) trace back to Severson (Nature Energy 2019), Attia (Nature 2020), and Zhu (Nature Communications 2022); reproducibility risk due to fabricated raw data is minimal.
No image-manipulation risk surface. The paper is an algorithmic study without Western blots, microscopy, or experimental imaging.
Statistical plausibility of Table IV. Reported runtimes (e.g., 1.88 s, 2.09 s, 1.29 s) and MAPE values (e.g., 3.81%, 5.16%) show normal decimal-level variation consistent with stochastic deep-learning training; no "too-clean" pattern observed.
Timeline consistency. MATLAB R2024a and references dated 2025 (e.g., [6], [12], [33]) are coherent with a late-2025 submission/publication.

Evidence highlights

Equation (15) appears at the end of Section III.C describing the compression rate.
Evaluation metric definitions in Section IV.A reuse numbers (7)–(9): "...Mean Absolute Percentage Error (MAPE)... (7) Relative Root Mean Square Error (RRMSE)... (8) R² ... (9)".
Section III.D states: "...the entire network is optimized using the Adam [20] optimizer..." with reference [20] listed as a 2025 IEEE TTE battery paper, not Kingma & Ba (2014).
Datasets map to Severson et al. (Nature Energy 2019), Attia et al. (Nature 2020), Zhu et al. (Nature Communications 2022).
Reported metric values in Table IV (e.g., MAPE 3.81%, 5.16%; test times 1.88 s, 2.09 s, 1.29 s) show natural variability.

Notes

No recommendation to file a formal misconduct allegation on the basis of current evidence.
A PubPeer comment highlighting the equation-numbering duplication and the Adam citation error is appropriate.
The authors should be encouraged to publish an Erratum correcting equation numbering in Section IV.A and replacing the Adam citation with Kingma & Ba (2014).
All findings are limited to what could be verified from the PDF text provided; the underlying source code and raw predictions were not independently inspected.