Low-cycle fatigue reliability analysis of curvic couplings on an artificial neural network surrogate model

Due to the structural complexity of curvic couplings, conducting extensive physical experiments poses considerable challenges. Consequently, an artificial neural network (ANN) surrogate model is proposed. Using both standard and notched specimens, a fatigue life model for low-cycle loading is formulated within the framework of critical plane theory. The maximum error in the Smith-Watson-Topper (SWT) parameter within a 1 mm region is constrained to 7.52%. A parameterized model is established to decompose the typical geometry of the curvic coupling structure, followed by finite element analysis to extract the SWT damage parameter. By using the created life model, the extracted SWT values are used to forecast the fatigue life of the root. The ANN surrogate model is trained with key structural parameters as inputs and the predicted coupling life as output. Employing quasi-Monte Carlo sampling, reliability analysis is conducted. Finally, a physical simulation specimen is designed and tested to verify the consistency of the SWT parameter within the critical distance. The experimental results lie within the ±2σ range of the predicted life distribution.