For over a century, researchers have been developing fatigue damage accumulation models based on purely empirical observations or based on theories derived thereof. Nonetheless, the Palmgren-Miner model published in 1945 remains the industry standard for fatigue life prediction of components subjected to variable amplitude loading. High-cycle fatigue is a complex multi-scale phenomenon governed by micro and macro-scale damage mechanics. Fundamental knowledge is lacking on how to link these damage mechanisms occurring at different scales. Physics-informed data-driven methods have recently become an attractive approach to model complex physical mechanisms, such as fatigue.

This project aims to develop a probabilistic physics-informed neural network for fatigue life prediction of components subjected to variable amplitude loading. A new methodology for microstructural characterization of fatigue damage will be developed and used to validate crystal plasticity models. These will in turn be used to study the microstructural origin of damage initiation. That information will be used to develop a microstructurally-informed fatigue damage accumulation model. Experimental data will be combined with model predictions to train the physics-informed neural network. Finally, a sensitivity analysis of the neural network features will be used to gain a deeper understanding of the fatigue damage accumulation mechanisms at different scales.