Non-hierarchical multi-output multi-fidelity Gaussian processes using a structure-aware composite kernel

초록

Multi-fidelity surrogate (MFS) modeling is crucial for alleviating computational burdens in engineering design. However, existing frameworks for multi-output systems often overlook the inter-output correlations inherent in low-fidelity (LF) sources. To address this limitation, this paper proposes the structure-aware multi-output multi-fidelity (S-MOMF) framework, which explicitly leverages these correlations. A linear model of coregionalization (LMC) is employed to characterize the output dependencies within each LF source. Subsequently, a composite kernel is formulated to incorporate this structural information into the high-fidelity (HF) model by leveraging the LF task covariance structure as a metric via the Mahalanobis distance. To ensure scalability and parameter efficiency, input fusion and parameterized task covariance strategies are integrated. Numerical benchmarks and a photolithography simulation in semiconductor manufacturing indicate that the proposed approach can improve predictive accuracy and robustness when informative LF inter-output covariance is available. © 2026 Elsevier B.V.

키워드