Integrity Report: "Prediction of Ground Settlement Induced by Shield Tunneling Based on Self-Attention–Recurrent Neural Network Model" (DOI: 10.16285/j.rsm.2024.0980)

Academic fraud report · Geng Detector

Summary

This is a translation/summary of a Chinese 'Geng' academic-integrity report flagging a paper in Rock and Soil Mechanics (岩土力学, DOI: 10.16285/j.rsm.2024.0980) by Song Muyuan and colleagues as highly suspicious. Six issues are documented. Most critically, Table 4 reports MSE and RMSE values that violate the mathematical identity RMSE = sqrt(MSE) for all four models (e.g., MSE=0.0015 vs RMSE=0.0348, since 0.0348^2 = 0.00121), strongly suggesting fabricated or patched-together data. Second, the abstract claims an optimal MAE of 0.0366 mm, while Table 4 and the conclusion give 0.0336 mm — a ~9% relative discrepancy on the headline result. Additional concerns include a figure caption mis-pointing readers to an LSTM cell diagram instead of the monitoring-point layout, two duplicate attention formulas (Eq. 17 and 18), suspiciously uniform 95% confidence-interval coverage values (95.18%–96.42%), and a reference dated 2018 that actually belongs to a 2023 volume. Verdict: highly suspicious pending verification of raw data and code; these issues together imply careless reporting at minimum, with non-negligible risk of data fabrication.

Verdict

Highly suspicious. Multiple independent red flags — including a fundamental algebraic inconsistency between reported MSE and RMSE values and an internal contradiction in the headline MAE — warrant formal investigation by the journal editors and, if possible, a request for raw data, code, and replication materials from the authors.

Key findings

Mathematical inconsistency between MSE and RMSE (Table 4 vs. Equations 23–24). RMSE must equal √MSE, but no row in Table 4 satisfies this identity.
Internal contradiction on the main reported MAE. Abstract states 0.0366 mm; Table 4 and Conclusion state 0.0336 mm.
Figure–text mismatch in Section 3.1.1. The text references Figure 1 for the monitoring-point layout, but Figure 1 depicts an LSTM unit structure; the actual layout figure is Figure 7.
Duplicated formula. Equations (17) and (18) are identical expressions for the attention scaling factor.
Suspiciously uniform 95% coverage values (Table 3). Six reported coverage rates fall within an unusually narrow 95.18%–96.42% band.
Reference [36] date error. A 2018 citation corresponds to *Engineering Structures* vol. 276, which was actually published in 2023.

Evidence highlights

MSE/RMSE mismatch (Table 4):
SAM-Bi-LSTM: MSE = 0.0015, RMSE = 0.0348 → 0.0348² = 0.00121104 ≠ 0.0015.
SAM-Bi-GRU: MSE = 0.0112, RMSE = 0.0992 → 0.0992² = 0.00984064 ≠ 0.0112.
Bi-LSTM: MSE = 0.0421, RMSE = 0.1910 → 0.1910² = 0.036481 ≠ 0.0421.
Bi-GRU: MSE = 0.0937, RMSE = 0.2840 → 0.2840² = 0.080656 ≠ 0.0937.
MAE contradiction: Abstract: 0.0366 mm; Table 4 and Conclusion §4(3): 0.0336 mm (relative difference ≈ 9%).
Figure mismatch: Section 3.1.1 cites Figure 1 for "6 settlement monitoring points"; Figure 1 is an LSTM unit diagram, Figure 7 is the actual tunnel/sampling-point layout.
Duplicate formula: Eq. (17) and Eq. (18) both equal α = softmax[QK^T / (d_k)^{1/2}].
Coverage values (Table 3): 95.83%, 96.42%, 95.18%, 95.23%, 95.45%, 95.18% — all within 95.18%–96.42%.
Reference [36]: FATHNEJAT et al., *Engineering Structures*, 2018, 276 → volume 276 corresponds to 2023.

Notes

The DOI of the paper under review is 10.16285/j.rsm.2024.0980.
All numeric values above are reproduced exactly as reported in the original Chinese integrity report and were not independently re-verified against the published PDF.
The MSE/RMSE paradox is the single most decisive issue: a correct evaluation pipeline cannot simultaneously produce these paired numbers. This strongly suggests either data fabrication, mixing of results from unrelated runs, or post-hoc manual editing of one column without recomputing the other.
The MAE discrepancy and the duplicate formula are consistent with sloppy writing/revision rather than deliberate misconduct, but they erode confidence in the reported results.
The uniform 95%-coverage values, while not impossible, are statistically unusual for real field-monitoring data and merit scrutiny.
Recommended actions (per the original report): post concerns on PubPeer, request raw data and code from the authors, and submit a formal complaint to the journal editor requesting verification of the core numerical results.
This report is AI-assisted and intended for academic discussion only; final determination of misconduct requires an official investigation.