A Hybrid Framework for Characterizing and Benchmarking Fatigue S-N Curves in Aluminum Alloys by Integrating Empirical and Data-Driven Approaches

The complicated and stochastic nature, coupled with uncertainties and data scatter, poses challenges in developing a general fatigue model. This study introduces a hybrid framework that integrates an empirical model with data-driven approaches, aiming to address data scarcity and streamline the fatigue characterization of aluminum alloys. It was found that support vector regression (SVR) and neural network (NN) exhibit superior performance, with SVR achieving a mean absolute error (MAE) of 0.13 (cycles to failure in log scale) for training and 0.14 for testing, and NN reaching an MAE of 0.15 for training and 0.16 for testing data. The employment of leave-one-group-out-cross-validation (LOGOCV) ensured the generalizability of the models, with the effectiveness confirmed by the actual-predicted life plot. The results demonstrated that almost 98% of predicted data fell within the life factor of +/- 1. This methodology reduces the requirement for experimentation and provides the prospect of automating fatigue design and characterization.