Neural Networks for Stress Intensity Factor Vertex Prediction

Abstract

Structural fatigue in aging metallic aircraft is a primary driver of sustainment costs for the U.S. Air Force, significantly impacting fleet readiness. Fatigue life prediction tools like AFGROW depend on interpolating between computationally expensive stress intensity factors (K-solutions) to approximate unknown values. However, interpolation errors in the current approach introduce uncertainty and force overly conservative maintenance schedules. This paper investigates the use of a machine learning surrogate to replace AFGROW’s dimensionreduction interpolation for the finite-width c orner-cracked hole geometry. We developed a robust data processing pipeline for a large FEA dataset and trained a neural network model. Our results reveal a critical insight: the surrogate model offers substantial performance gains over AFGROW’s interpolation in low-data regimes, emphasizing both the potential of the model and its sensitivity to dataset size. For the original, sparse dataset—which is characteristic of computationally expensive problems—the neural network significantly o utperformed the baseline interpolation, reducing the mean absolute percentage error (MAPE) by over 40% (from 2.77% to 1.60%) and achieving an R² value exceeding 0.99. However, experiments on synthetically generated dense datasets showed that the traditional interpolation method becomes more accurate as the data grid becomes less sparse. This study concludes that while neural network surrogates offer a powerful, high-fidelity solution for computationally intensive engineering problems, their adoption should be guided by a careful analysis of data density after dataset has been cleaned of outliers. It also highlights the necessity of employing rigorous, application-relevant validation strategies that move beyond simplistic random splits to accurately assess model performance.