Physics-guided neural network for fatigue life prediction of FCC-based multi-principal element alloys

• Published in: Scripta materialia 2024-12, Vol.253, p.116307, Article 116307
• Main Authors: Ren, Jingli, Xiao, Lu
• Format: Article
• Language: English
• Subjects: Fatigue life prediction; Multi-principal element alloys; Physics-guided neural network
• ISSN: 1359-6462
• DOI: 10.1016/j.scriptamat.2024.116307

We propose a physics-guided neural network (PGNN) model to predict the fatigue life of multi-principal element alloys (MPEA). It is found that the knowledge mined by purely data-driven methods violates well-known physical laws, exhibiting a negative correlation between ductility and fatigue life. To address this physical inconsistency, a PGNN prediction model is developed by transforming physical knowledge into constraints on the neural network's weights. The SHAP and univariate analysis conducted by the PGNN model show physically consistent results, indicating a positive correlation between yield strength and ductility with fatigue life. Moreover, the proposed PGNN model outperforms solely data-driven models, with higher goodness of fit (R2 = 86.5%), smaller error (MSE = 0.125), and better generalization ability. This work provides a fast and low-cost model to predict the fatigue life of FCC-based MPEA, and the proposed PGNN model can overcome the limitations of purely data-driven models and produce accurate and physically consistent results.