Physics-informed AI predicts steel fatigue life from micrographs in 65ms

Evaluating structural steel fatigue life traditionally requires mechanical testing that spans tens to hundreds of hours, a major bottleneck for rapid quality control. In a new arXiv paper, researcher Aryuemaan Kumar Chowdhury presents CV, a computer vision framework that sidesteps physical testing entirely. The pipeline combines a seven-stage OpenCV preprocessing routine, a 28-dimensional physics-informed feature extractor quantifying crack morphology, grain structure, porosity, and texture, and a CNN regression head trained with a Gaussian negative log-likelihood (GNLL) loss to jointly predict fatigue life and sample-specific uncertainty.

On a synthetic micrograph benchmark, the team compared three architectures—SE-CNN, ResNet-50, and VGG-16. ResNet-50 emerged as the winner, achieving an R² of 0.93, an RMSE of 0.18 log-cycles, and a macro-F1 score of 0.91. The GNLL loss also dramatically improved model calibration, reducing Expected Calibration Error by 76% compared to a mean-squared-error baseline (ECE dropped from 0.089 to 0.021). Grad-CAM analysis confirmed the model attends to metallurgically meaningful microstructural features, validating that it learns relevant physics rather than spurious patterns. The entire pipeline runs in under 65 ms per image, and both the code and synthetic dataset generator are open-sourced.

While these results are promising, the study relies entirely on synthetic micrographs—so the performance represents methodological soundness under simulated conditions. The author explicitly flags a domain-transfer study on real field samples as the immediate next step. Still, the speed and accuracy suggest that AI-driven micrograph analysis could soon replace or augment destructive fatigue testing in manufacturing, enabling real-time quality control and safer material design. If real-world validation succeeds, CV could become a standard inspection tool for steel producers and structural engineers.