A hybrid physics-informed neural network for predicting fatigue life of axially loaded grouted connections for offshore wind turbine structures

Zhao, Siyu and Huang, Zhenyu and Zhou, Yingwu and Xing, Feng and Hou, Xiaonan and Ye, Jianqiao (2026) A hybrid physics-informed neural network for predicting fatigue life of axially loaded grouted connections for offshore wind turbine structures. Engineering Structures, 351: 122043. ISSN 0141-0296

Text (ES)
ES.pdf - Accepted Version
Available under License Creative Commons Attribution.
Download (5MB)

Abstract

As a critical component of an offshore wind turbine foundation, grouted connections are susceptible to fatigue failure under long-term loading conditions, which makes the assessment of their fatigue behaviour essential for the overall structural integrity and maintenance. Compared to costly experimental characterization and computationally expensive numerical modelling, there is a high demand for developing a fast, efficient and accurate surrogate model to predict fatigue life. This paper develops a novel hybrid Physics-Informed Neural Network (PINN) model that integrates both simplified physical constraints and hidden physical laws to predict the fatigue life of axially loaded grouted connections, where the physical knowledge is the relationship between fatigue life and fatigue-related parameters. The results show that the developed hybrid PINN model achieves superior prediction accuracy compared to the current codes of practice, the conventional Deep Neural Network (DNN) model, the PINN model integrating simplified physical constraints (S-PINN), and the PINN model integrating hidden physical laws (H-PINN). To enhance the interpretability of the model, Shapley Additive Explanations (SHAP) analysis and physical consistency analysis are conducted to assess the contribution of each feature to the fatigue life and to investigate the distribution of predictions with respect to physical consistency. It’s found that the new hybrid PINN model produces predictions that exhibit a higher degree of physical consistency than the purely data-driven DNN model, demonstrating the reliability and robustness of the model.