A physics-informed bayesian framework for fatigue crack growth under multiple conditions with reduced experimental Effor: Application to the NH3 tank material

This study develops an integrated experimental–computational framework for characterising fatigue crack growth behaviour across different crack orientations, temperatures, and load ratios while minimising experimental demand by employing a Physics-Informed Bayesian Neural Network (PIBNN) that enables cross-orientation information sharing. While extensive research has focused on the effects of temperature, load ratio, and residual stress, the influence of crack orientation has received comparatively limited attention. In structures, cracks can propagate along different orientations, leading to variations in residual stress fields and crack-closure effects that influence crack-growth rates. The study systematically analyses the influencing factors governing fatigue crack growth across different orientations. Conventional experimental approaches require repeated testing for each combination of crack orientation, temperature, and load ratio, resulting in substantial experimental cost. The proposed PIBNN model addresses this limitation by enabling cross-condition knowledge transfer across crack orientations, temperature levels, and load conditions. The Bayesian framework explicitly quantifies crack-growth uncertainty using cross-condition data, thereby reducing the number of experiments required to establish reliable design curves. The framework is applied to FH36 steel relevant to low-temperature ammonia and CO₂ transport vessels. The results demonstrate that the method accurately predicts crack growth behaviour while reducing testing requirements, providing a reliable and efficient probabilistic framework for crack-growth assessment under multiple conditions.