Predicting Steel Fatigue Life from Micrographs Using Physics-Informed Deep Learning

Technical Summary: FATIGUECV – Predicting Steel Fatigue Life from Micrographs Using Physics-Informed Deep Learning

Problem Statement
Conventional evaluation of structural steel fatigue life ($N_f$) relies on destructive, load-controlled S–N tests that require 50 to 500 hours per specimen. This timescale renders the method impractical for rapid quality control in lightweight-critical sectors (automotive, rail, aerospace). While optical microscopy is a routine quality control step that captures microstructural drivers of fatigue (grain structure, crack morphology, porosity), there is currently no established method to convert this qualitative visual data into a quantitative, calibrated estimate of fatigue life. Furthermore, existing machine learning approaches for fatigue typically rely on tabular inputs (composition, mechanical properties) or fail to provide per-sample uncertainty quantification, which is critical for safety assurance.

Methodology: The FATIGUECV Pipeline
The author presents FATIGUECV, an end-to-end computer vision framework that estimates $\log_{10}(N_f)$ and sample-specific uncertainty directly from optical micrographs. The pipeline consists of four sequential layers:

1. Synthetic Data Generation:
   Due to the lack of public datasets pairing micrographs with ground-truth fatigue life, the author developed a physics-constrained synthetic generator. It simulates micrographs using Voronoi tessellation for grain structures, correlated random walks for crack propagation, and Poisson-distributed voids. Labels are assigned via a physics-motivated regression equation (Eq. 1) that incorporates crack severity, porosity, grain size (Hall–Petch relation), and inclusion density, plus inherent material scatter.

2. Seven-Stage Preprocessing (OpenCV):
   A specialized pipeline cleans raw micrographs to remove artifacts common in optical metallography. Stages include:

   • Quality control gating (Laplacian blur, SNR).
   • Border cropping to remove scale bars.
   • Non-local means denoising (preserving edge sharpness).
   • Rolling-ball background subtraction and CLAHE for illumination correction.
   • Multi-orientation black-hat morphology and Scharr-gradient fusion for crack enhancement.
   • Sharpening and resizing to $224 \times 224$ px.
   • Result: This stage increases crack detection recall from 61% to 89% and improves illumination uniformity by 54%.

3. Physics-Informed Feature Extraction:
   Each image is mapped to a 28-dimensional feature vector ($x \in \mathbb{R}^{28}$) explicitly derived from fatigue mechanics. Features are grouped into five categories: crack morphology (count, length, branching), grain structure (size, distribution), porosity (void fraction), texture (GLCM metrics), and gradient/fractal properties (fractal dimension $D_f$).

4. Dual-Head CNN Regression:
   The system employs a Convolutional Neural Network (tested with SE-CNN, ResNet-50, and VGG-16 backbones) with two output heads:

   • Mean Head ($\hat{\mu}$): Predicts $\log_{10}(N_f)$.
   • Log-Variance Head ($\hat{s}$): Predicts $\log \hat{\sigma}^2$, encoding per-sample aleatoric uncertainty.
   • Training Objective: The network is trained using a Gaussian Negative Log-Likelihood (GNLL) loss. This allows the model to learn heteroscedastic uncertainty, where the predicted variance adapts to the input (e.g., higher uncertainty for heavily damaged specimens with high scatter).

Key Results
Evaluated on a held-out synthetic test set (200 images):

• Regression Performance: The ResNet-50 architecture achieved the best performance with an $R^2 = 0.93$, RMSE = 0.18 log-cycles, and MAE = 0.14. This outperformed the SE-CNN ($R^2=0.88$) and VGG-16 ($R^2=0.91$), as well as a tabular Support Vector Regression baseline ($R^2=0.87$) from prior literature.
• Uncertainty Calibration: The GNLL objective significantly improved calibration compared to a Mean Squared Error (MSE) baseline. The Expected Calibration Error (ECE) dropped from 0.089 (MSE) to 0.021 (GNLL), a 76% reduction. The nominal 95% confidence intervals empirically contained 93.8% of test samples.
• Risk Classification: The system classifies samples into four risk tiers (LOW to CRITICAL) with a macro-F1 score of 0.91. Critical-tier recall was 0.90.
• Explainability: Grad-CAM analysis confirmed that the network attends to metallurgically meaningful features. Attention maps shifted from grain boundaries (LOW risk) to crack nucleation sites (MEDIUM), crack tips (HIGH), and pervasive crack networks (CRITICAL), aligning with established fatigue damage stages (Suresh, 1990).
• Efficiency: Inference time is under 65 ms per image.

Significance and Claims
The paper positions FATIGUECV as a methodological proof of concept rather than a deployable solution for real-world steel immediately. The author explicitly states that all results are derived from a synthetic benchmark where the ground truth is defined by the simulation physics.

The primary contributions are:

1. First Image-to-Life Regression with Uncertainty: It is the first system to jointly perform end-to-end regression from micrograph to fatigue life while providing calibrated per-sample uncertainty bounds.
2. Physics-Guided Inductive Bias: The integration of a 28-dimensional feature vector derived from fatigue mechanics and a GNLL loss function creates a strong inductive bias that recovers known physics relationships from images.
3. Metallography-Specific Preprocessing: A novel seven-stage pipeline tailored specifically to optical microscopy artifacts, significantly improving feature visibility.
4. Open-Source Synthetic Benchmark: The release of a configurable, physics-labeled synthetic dataset generator to enable reproducible benchmarking and future domain-transfer studies.

Limitations and Future Work
The author is modest regarding the current scope, highlighting four key limitations:

1. Synthetic-to-Real Gap: The model has not been tested on real steel micrographs; domain adaptation (e.g., style transfer) and re-validation on physical specimens are required.
2. Material Scope: The pipeline is tuned for carbon and low-alloy steels; other alloys (Ti, Al) would require redesign.
3. 2-D Projection: Optical microscopy cannot detect subsurface defects; integration with 3-D modalities (X-ray CT) is planned.
4. Holistic Prediction: The model predicts total life ($N_f$) rather than separating initiation and propagation phases.

The paper concludes that while the method demonstrates soundness under simulated conditions, the immediate next step is a domain-transfer study on real field samples to validate predictive validity.