Physics-based machine learning for fatigue lifetime prediction under non-uniform loading scenarios
Abedulgader Baktheer, Fadi Aldakheel
TL;DR
This work develops a physics‑based machine learning surrogate (φML) to predict fatigue lifetime under non‑uniform loading. By training a feed‑forward neural network on data generated from an anisotropic continuum damage model and enforcing physics through custom losses, φML achieves accurate predictions with limited experimental data and dramatically reduced computation time. The approach generalizes from two‑level to multi‑level loading scenarios, aligning with experimental trends that Palmgren‑Miner fails to capture, and enabling real‑time fatigue assessment within digital twins. The results demonstrate substantial speedups, robustness to data scarcity, and potential for integrating advanced multiscale fatigue phenomena into engineering decision workflows. Overall, φML offers a practical, physics‑informed route to reliable fatigue life prediction for complex loading histories in concrete and related materials.
Abstract
Accurate lifetime prediction of structures subjected to cyclic loading is vital, especially in scenarios involving non-uniform loading histories where load sequencing critically influences structural durability. Addressing this complexity requires advanced modeling approaches capable of capturing the intricate relationship between loading sequences and fatigue lifetime. Traditional fatigue simulations are computationally prohibitive, necessitating more efficient methods. This study highlights the potential of physics-based machine learning ($φ$ML) to predict the fatigue lifetime of materials. Specifically, a FFNN is designed to embed physical constraints from experimental evidence directly into its architecture to enhance prediction accuracy. It is trained using numerical simulations generated by a physically based anisotropic continuum damage fatigue model. The model is calibrated and validated against experimental fatigue data of concrete cylinder specimens tested in uniaxial compression. The proposed approach demonstrates superior accuracy compared to purely data-driven neural networks, particularly in situations with limited training data, achieving realistic predictions of damage accumulation. Thus, a general algorithm is developed and successfully applied to predict fatigue lifetimes under complex loading scenarios with multiple loading ranges. Hereby, the $φ$ML model serves as a surrogate to capture damage evolution across load transitions. The $φ$ML based algorithm is subsequently employed to investigate the influence of multiple loading transitions on accumulated fatigue life, and its predictions align with trends observed in recent experimental studies. This work demonstrates $φ$ML as a promising technique for efficient and reliable fatigue life prediction in engineering structures, with possible integration into digital twin models for real-time assessment.