Integrity Review Report: 'SiMBA-Augmented Physics-Informed Neural Networks for Industrial Remaining Useful Life Prediction' (DOI: 10.3390/machines13060452)

Academic fraud report · Geng Detector

Summary

This report flags the paper 'SiMBA-Augmented Physics-Informed Neural Networks for Industrial Remaining Useful Life Prediction' (Machines, 2025; DOI: 10.3390/machines13060452) as highly suspicious based on textual, tabular, and methodological cross-checking. Four major issues are identified. First, in Table 4 the proposed model's RMSE values across four C-MAPSS subsets (FD001–FD004) cluster suspiciously tightly at 16.94, 16.91, 16.92, and 17.45—an implausibly narrow spread given the datasets' very different operating conditions and fault modes. Second, the Score value of 1665 for FD002 is identical between the proposed model and the comparator Cau-AttnPINN (Ref. [27]), whose first author is also Min Li, suggesting a possible copy-paste error. Third, the paper's core model name is inconsistently spelled: 'SiMBA' in the title and abstract versus 'SiMAB' throughout the data tables. Fourth, the submission-to-acceptance timeline (36 days, with only 2 days between revision and acceptance) and the claimed CPU-only laptop environment are inconsistent with the reported experimental workload. Verdict: highly suspicious; verification of raw logs, code, and peer-review records is recommended.

Verdict

🟠 Highly Suspicious. Multiple independent red flags—implausibly uniform benchmark scores, an identical Score value matching a prior paper by the same first author, a persistently misspelled core model name, an unrealistic review timeline, and a questionable compute-environment claim—converge to suggest possible data manipulation, copy-paste reuse, and inadequate peer review. No original high-resolution figures were analyzed; all findings derive from the published PDF text and tables.

Key findings

Suspiciously uniform RMSE across heterogeneous datasets (Table 4). The proposed model reports RMSE of 16.94 (FD001), 16.91 (FD002), 16.92 (FD003), and 17.45 (FD004). Given that FD001/FD003 contain a single operating condition while FD002/FD004 contain six, near-identical generalization error is statistically implausible and suggestive of hyperparameter/seed tuning.
Identical Score value with a prior paper by the same first author (Tables 4 & 5). The proposed model's Score on FD002 (1665) exactly matches that of Cau-AttnPINN (Ref. [27], first author: Min Li). Because the NASA Score function is a nonlinear asymmetric penalty, an exact integer match between two distinct models is extremely unlikely and consistent with accidental data reuse.
Persistent misspelling of the core model name. The text uses SiMBA (Simplified Mamba), while Tables 2, 3, 4, and 5 systematically use SiMAB, indicating an uncontrolled find-and-replace failure rather than a typo.
Implausible review timeline. Received 17 April 2025; Revised 21 May 2025; Accepted 23 May 2025. The revision-to-acceptance interval is 2 days, incompatible with a paper claiming extensive ablations and complex model training.
Compute-environment contradiction. Section 2.5 states experiments ran on a Xiaomi laptop with 12th-gen i5-12450H CPU and 16 GB RAM with no GPU mentioned, yet the model incorporates FFT, bidirectional SSM, and attention modules trained for up to 300 epochs on long C-MAPSS sequences—workloads that would take months on CPU alone.
FLOPs unit error (Table 5). Reported FLOPs of 1728 (AttnPINN) and 5790 (proposed model) are implausibly small; the values almost certainly omit an MFLOPs/GFLOPs multiplier, indicating unfamiliarity with standard reporting conventions.

Evidence highlights

Table 4 (p. 12): Proposed-model RMSE = 16.94 / 16.91 / 16.92 / 17.45 across FD001–FD004; Score on FD002 = 1665 (identical to Cau-AttnPINN, Ref. [27]).
Tables 2–5: Headers and leftmost columns consistently labeled "SiMAB-PINN" while the title/abstract use "SiMBA."
Table 5 (pp. 13–14): FLOPs listed as raw integers 1728 and 5790, inconsistent with standard FLOPs/MFLOPs/GFLOPs reporting.
Header / Section 2.5: Received 17 Apr 2025; Revised 21 May 2025; Accepted 23 May 2025; hardware = Xiaomi laptop, i5-12450H, 16 GB RAM, no GPU.

Notes

DOI preserved exactly: 10.3390/machines13060452.
All numeric evidence above is taken verbatim from the published PDF; no findings have been invented.
Limitations: original high-resolution figures, raw training logs, and source code were not available for inspection. Findings should therefore be treated as strong indicators warranting institutional investigation rather than a definitive determination of misconduct.
Recommended actions: (1) request raw loss curves, random seeds, wall-clock training times, and code from the corresponding author; (2) raise concerns on PubPeer focusing on the RMSE uniformity, the 1665 Score coincidence, and the SiMAB/SiMBA inconsistency; (3) notify the editorial office of *Machines* about the review-timeline anomaly; (4) audit other recent MDPI publications by the same group for similar patterns.