Predicting Steel Fatigue Life from Micrographs Using Physics-Informed Deep Learning
Physics-informed deep learning framework predicts steel fatigue life (log N_f) directly from optical micrographs, eliminating hours of mechanical testing 28-dimensional feature extractor quantifies crack morphology, grain structure, porosity, and texture with metallurgical grounding ResNet-50 achieves R² = 0.93, RMSE = 0.18 log-cycles, macro-F1 = 0.91 on synthetic micrograph benchmark Gaussian negative log-likelihood loss reduces Expected Calibration Error by 76% (ECE: 0.089 → 0.021) vs MSE base
Analysis
TL;DR
- Physics-informed deep learning framework predicts steel fatigue life (log N_f) directly from optical micrographs, eliminating hours of mechanical testing
- 28-dimensional feature extractor quantifies crack morphology, grain structure, porosity, and texture with metallurgical grounding
- ResNet-50 achieves R² = 0.93, RMSE = 0.18 log-cycles, macro-F1 = 0.91 on synthetic micrograph benchmark
- Gaussian negative log-likelihood loss reduces Expected Calibration Error by 76% (ECE: 0.089 → 0.021) vs MSE baseline
- Open-sourced pipeline processes images in under 65ms; domain transfer to real field samples is the immediate next step
Why It Matters
This work demonstrates how physics-informed deep learning can bridge materials science and computer vision for rapid quality control in structural engineering. The uncertainty-aware predictions (sample-specific σ̂) enable risk-aware decision-making in fatigue-critical applications where overconfident errors could be catastrophic.
Technical Details
- Pipeline: Seven-stage OpenCV preprocessing removes artifacts, followed by a 28-dimensional physics-informed feature extractor quantifying crack morphology, grain structure, porosity, and texture
- Model: CNN regression with Gaussian negative log-likelihood (GNLL) loss jointly predicts log N_f and sample-specific uncertainty σ̂
- Architectures evaluated: SE-CNN, ResNet-50, VGG-16 on synthetic micrograph benchmark
- Performance: ResNet-50 achieves R² = 0.93, RMSE = 0.18 log-cycles, macro-F1 = 0.91; Grad-CAM confirms attention to metallurgically meaningful microstructures
- Calibration: GNLL objective reduces Expected Calibration Error by 76% compared to MSE baseline (ECE: 0.089 → 0.021)
Industry Insight
- Accelerated testing pipelines: Sub-65ms inference enables real-time fatigue life screening in manufacturing quality control, potentially replacing days of mechanical testing
- Uncertainty quantification as safety critical: Sample-specific uncertainty estimates (σ̂) allow engineers to flag low-confidence predictions for manual review, crucial for fatigue-critical infrastructure
- Synthetic-to-real domain gap: While results are methodologically sound on synthetic data, the immediate next step—domain transfer to real field samples—will determine practical viability; practitioners should monitor this transition closely before adoption
Disclaimer: The above content is generated by AI and is for reference only.