Data Fusion of Multi-fidelity Systems via Latent Variable Gaussian Process for Active Learning Applications
Date:
In this work, we develop a new multi-fidelity modeling method based on the Latent Variable Gaussian Process (LVGP), a Gaussian Process (GP) based approach that can accommodate mixedvariable (quantitative and qualitative) variables as predictors. LVGP can straightforwardly handle the multi-fidelity modeling problem by representing the fidelity level of different models as a qualitative variable. The LVGP learns the correlations between fidelity models through a Gaussian kernel and represents them with latent variables. Unlike the sequential training architecture, the LVGP only requires one training to construct a single hypersurface that simultaneously accommodates all fidelity models and their correlations. This representation enables the direct conditioning of high-fidelity model predictions on all low-fidelity models, minimizing the dilution of information and uncertainty propagation.
