Physics of Failure Transfer Learning for Aerospace Electronic Component Considering Manufacturing Parameter

The reliability of aerospace electronic components operating under extreme conditions is crucial to mission success. Physics-of-failure (PoF) modeling provides mechanism-based reliability analysis, yet its deployment in aerospace electronics is hindered by scarce target-domain data, high test cost, and strong sensitivity of PoF coefficients to manufacturing variability. This article proposes a PoF-guided transfer learning framework that quantifies the influence of manufacturing parameters on PoF model coefficients and evaluates source–target transferability using a dispersion-aware maximum mean discrepancy (MMD) metric that jointly accounts for mean discrepancy and distance variance in reproducing kernel Hilbert space (RKHS). A physically consistent transfer model is then constructed via a hybrid base learner (HBL) with an RMSE-driven reweighting strategy. Experiments on MOSFET and aerospace electromagnetic relays show that domain similarity grades align well with transfer performance, achieving a perfect Spearman’s rank correlate ion coefficient (Ρ = 1) in 83.3% (10/12) of evaluations. The proposed framework attains maximum relative errors of 6.08% and 2.19% for representative PoF coefficients, and achieves a minimum R² of 0.992 with a maximum mean squared error (mse) of 6.51 x 10^-5, outperforming individual tree-based learners and representative transfer baselines. Because the workflow only requires measurable manufacturing parameters, a parametric PoF form, and limited target-domain samples, it provides a generally applicable pathway for reliability-oriented PoF modeling and prediction of aerospace electronic components under manufacturing variations.