Integrity review: Constructing boundary-identical microstructures via guided diffusion for fast multiscale topology optimization (DOI: 10.1016/j.cma.2025.117735)

Academic fraud report · Geng Detector

Summary

Verdict: No substantive academic fraud detected. The paper, authored by Jingxuan Feng et al. and published in Computer Methods in Applied Mechanics and Engineering (2025), passed textual, numerical, and logical consistency checks. The central ablation figures were reverse-verified against Tables A.1 and A.2 with exact agreement (R² mean drop of 0.0516 and DoM mean drop of 0.0255 after removing self-conditioning), demonstrating tight internal consistency rather than fabricated numbers. The methodology, computational setup (8× RTX 3090 GPUs, 1000 epochs), and code/data availability statements are consistent with contemporary deep-learning practice (DDPM, classifier-free guidance, U-Net). Two minor textual typos were noted: '0.7000' likely should read '0.070', and '0.8800' should read '0.088', as supported by Table A.2 values (0.991 − 0.921 = 0.070). Limitations include the absence of pixel-level image forensics on figures and reliance on textual analysis only.

Verdict

Clean / No fraud indicators. The work shows high internal numerical consistency and coherent methodology. Only minor typographical errors were identified.

Key findings

Numerical self-consistency (positive verification): Reverse-calculation of ablation results matches the prose exactly. Table A.1: Baseline R² mean 0.9653 vs. w/o self-conditioning 0.9137 → Δ = 0.0516, identical to the paper's claim of "an average decrease of 0.0516 in R²-score." Table A.2: Baseline DoM mean 0.99175 vs. 0.96625 → Δ = 0.0255, matching the stated reduction.
Probable typographical errors: The prose states "a maximum reduction of 0.7000 (Table A.2)" and "maximum reductions of 0.8800." Table A.2 shows the largest DoM difference in column 11 as 0.991 − 0.921 = 0.070, suggesting 0.7000 should be 0.070. Likewise, 0.8800 should read 0.088 based on Table A.1 low-level-layer differences. These appear to be decimal-point slips rather than fabricated claims.
Methodological plausibility: Tech stack (DDPM, self-conditioning, U-Net, classifier-free guidance) aligns with 2021–2022 advances; 8 × RTX 3090 GPUs training for 1000 epochs is reasonable for a diffusion model of this scale; claimed open-source release of code and a ~10⁵-scale dataset is consistent with practices in the field.
Timeline consistency: Submission in September 2024 with online publication on 14 January 2025 is internally coherent.

Evidence highlights

Table A.1: Baseline R² mean = 0.9653; w/o self-conditioning R² mean = 0.9137; Δ = 0.0516 ✓
Table A.2: Baseline DoM mean = 0.99175; w/o self-conditioning DoM mean = 0.96625; Δ = 0.0255 ✓
Table A.2, column 11: 0.991 − 0.921 = 0.070 (text says "0.7000") — likely a typo
Table A.1 low-level layer maximum Δ ≈ 0.088 (text says "0.8800") — likely a typo
DOI: 10.1016/j.cma.2025.117735; Journal: Computer Methods in Applied Mechanics and Engineering (CMAME), 2025

Notes

Pixel-level image forensics (e.g., duplicate regions, compression-inconsistent regions, western-blot-style splicing checks) were not performed because only text and table descriptions were available; this is the main limitation of the review.
The two decimal-point typos (0.7000 vs 0.070; 0.8800 vs 0.088) do not affect the paper's qualitative conclusions or the validity of the experiments, but a corrigendum would be appropriate.
No authorship anomalies, reference irregularities, or citation-pattern concerns were flagged from the supplied text.