From Design of Experiments to Neural Network Models for Predicting Sintered Nanosilver Joints Degradation Under Thermal Shocks

Abstract Sintered materials exhibit complex mechanical behaviors under cyclic loading, necessitating robust models for predicting responses such as dissipated energy, cyclic porosity damage, and stress level. This study integrates design of experiments (DoE) alongside neural network approaches to elucidate factor influences and enable accurate predictive modeling. An optimal Latin hypercube sampling in the DoE was used to systematically examine the effects of key factors (initial joint porosity, joint thickness, and maximum loading temperature) and their interactions on the responses, revealing dominant nonlinear relationships that inform material optimization. Leveraging the DoE-generated dataset, feedforward neural networks (FFNNs) and radial basis function networks (RBFNs) were compared in multi-output and single-output configurations. The multi-output FFNN, optimized with two hidden layers and a balanced training/validation/testing split, underperformed, particularly in mapping cyclic porosity change, highlighting challenges in joint prediction. Switching to three separate single-output FFNNs markedly improved accuracy but required optimizing more parameters, escalating computational demands. In contrast, RBFNs excelled, with both single- and multi-output variants yielding strong results. The multi-output RBFN proved the optimal compromise, achieving competitive mean squared error (MSE) with far fewer parameters than the single-output FFNNs, while minimizing marginal gains from disaggregated models. Using the multi-output RBFN, joint response mappings were finally successfully represented via contour plots, offering insights into factor variations for enhanced sintered material design. These findings underscore RBFNs' efficiency for multi-output tasks and joint properties optimization under thermal fatigue.