Abstract
This study presents a regularized Deep Kernel Learning (DKL) model for reliability analysis with high-dimensional uncertainties. Combining Deep Learning (DL) and Gaussian Process Regression (GPR), the model leverages DL for feature extraction and GPR for predictive variance estimation. This integration addresses the curse of dimensionality by mapping high-dimensional inputs to a lower-dimensional space for GPR, enhancing regression tasks. To combat overfitting, a prevalent issue in surrogate models with limited data, the model incorporates Dropout and L2 regularization techniques, making it more robust. An verification on two case studies demonstrates that regularized DKL surpasses both non-regularized DKL and GPR in predicting failure probabilities in high-dimensional scenarios.