AI Predicts Steel Fatigue Life from Micrographs Rapidly

Traditionally, evaluating the fatigue life of structural steels is a time-consuming process, often requiring tens to hundreds of hours of mechanical testing, which makes rapid quality control impractical. This paper introduces a novel computer vision framework, named CV, designed to quickly estimate the fatigue life (log Nf) of lightweight alloy steels directly from their optical micrographs. The CV pipeline involves a seven-stage OpenCV preprocessing routine to clean up image artifacts. Following this, a 28-dimensional physics-informed feature extractor quantifies critical microstructural characteristics such as crack morphology, grain structure, porosity, and texture. Finally, a Convolutional Neural Network (CNN) regression model is trained using a Gaussian negative log-likelihood (GNLL) loss function to simultaneously predict the fatigue life and its associated uncertainty. Evaluated on a synthetic micrograph benchmark, a ResNet-50 architecture within CV achieved impressive results, including an R-squared of 0.93 and a Root Mean Squared Error of 0.18 log-cycles. The GNLL objective significantly reduced the Expected Calibration Error by 76% compared to a standard mean-squared-error baseline, indicating better uncertainty quantification. Grad-CAM visualizations confirmed that the network focuses on metallurgically relevant features. The entire pipeline runs in under 65 milliseconds per image, and both the pipeline and synthetic dataset generator are open-sourced.