Crack retardation, commonly observed in metals under overload, helps extend fatigue life. Conventional methods for modeling behavior under variable loading are labor‐intensive and sensitive to measurement errors because they rely on cycle‐by‐cycle crack length measurements, plastic zone estimation, and stress intensity factor range (Δ K ) calculations. In this study, a deep learning approach based on crack‐tip strain fields, measured via digital image correlation, is proposed to evaluate overload effects on fatigue parameters, such as Δ K and fatigue crack growth rate (d a /d N ). For each data point, strain fields from successive load cycles are stacked to capture the evolution of crack‐tip behavior while reducing random noise. VGGNet16 and ResNet50 are used for classification and regression tasks. The classification task distinguishes three states: normal, overload, and recovery. Both convolutional neural networks (CNNs) achieve 99.4% accuracy in classification; VGGNet16 and ResNet50 achieve R 2 values of 0.96 and 0.89, respectively, for Δ K prediction and 0.92 and 0.91, respectively, for d a /d N prediction. The total evaluation time is on the order of seconds per data point, compared with hours for Wheeler/Elber analyses and days for finite element analysis. These results demonstrate the feasibility of CNNs in identifying overload‐induced crack states and predicting fatigue parameters directly from strain fields.
Choi et al. (Tue,) studied this question.