Accelerating the Measurement of Fatigue Crack Growth with Incremental Information-Based Machine Learning Approach

Measuring the fatigue crack growth rate via the crack growth experiment (<i>a</i>-<i>N</i> curve) is labor-intensive and time-consuming. A machine learning interpolation-extrapolation strategy (MLIES) aimed at enhancing the prediction accuracy of small-data models has been proposed to accelerate fatigue testing. Two specific approaches are designed by transforming <i>a</i>-<i>N</i> curve data from <i>N</i> to <i>ΔN</i> and from <i>a</i> to <i>Δa</i> (S1)/<i>Δa</i>/<i>ΔN</i> (S2) to enrich the data volume and leverage the incremental information. Thus, a simple and fast-responding single-layer neural network model can be trained based on the early-stage data points from fatigue testing and accurately predict the remaining part of an <i>a</i>-<i>N</i> curve, thereby enhancing the experimental efficiency. Through exponential data expansion and data augmentation, the trained neural network model is able to learn the underlying rules governing crack growth directly from the experimental data, requiring no explicit analytical crack growth laws. The proposed MLIES was validated on fatigue tests for aluminum alloy and titanium alloy samples under different experimental parameters. Results demonstrate its effectiveness in reducing testing time/cost by 15-32% while achieving over 30% higher prediction accuracy for the <i>a</i>-<i>N</i> curve compared to a traditional machine learning modeling approach. Our research offers a data-driven recipe for accurate crack growth prediction and accelerated fatigue testing.