Abstract
Abstract
High-fidelity simulation of nanoscale electro-thermal behavior in RF/microwave transmit/receive (T/R) modules increasingly relies on surrogate (metamodel) predictors to achieve tractable computational performance. When these metamodels are constructed in high-dimensional design spaces with partially misspecified covariance kernels, conventional Gaussian process (GP) and stochastic kriging approaches can exhibit unstable variance behavior, biased thermal predictions, and poor convergence diagnostics. This paper introduces a Hierarchical Heteroscedastic Gaussian Process (H2GP) framework with Iterative Diagnostic Feedback (IDF), designed to stabilize and accelerate sequential surrogate updating for multivariate electro-thermal analysis in RF T/R modules. The proposed architecture decomposes the global surrogate into hierarchically coupled sub-models that separately manage the mean response, local variance inflation, and adaptive kernel correction. Heteroscedastic simulation noise—arising from solver tolerances and mesh adaptivity—is explicitly modeled and fed back into the primary predictor. To address high-dimensional kernel misspecification, we construct an IDF loss that augments the negative log-marginal likelihood with penalties for miscalibrated predictive intervals and structured residuals, together with regularization on hyperparameters. We validate the framework against standard stochastic kriging and heteroscedastic GP baselines using a physics-based synthetic benchmark tailored to RF electro-thermal behavior. The synthetic experiments demonstrate that the proposed H2GP + IDF methodology substantially improves predictive coverage and hotspot localization under severe kernel misspecification, with only modest increases in interval width. These results highlight the importance of explicitly targeting predictive interval calibration—not only point prediction error—when building simulation metamodels for reliability-driven analysis in complex RF systems. The focus of this work is methodological; application to full-scale hardware testbeds and production EM-thermal workflows is identified as future work.