Fig. 3.

Linear (AR1) vs. non-linear (NARGP) Gaussian process regression: (a) Exact low- (red) and high-fidelity (blue) functions along with the observations used for training the multi-fidelity GP models (14 blue points, 50 red points). (b) Exact solution vs. the NARGP posterior mean and uncertainty. (c) AR1 predictions and its uncertainty vs. exact solution. (d) Cross-correlation structure between the exact low- and high-fidelity signals vs. the cross-correlation learnt by the NARGP and AR1 schemes trained on the given multi-fidelity dataset. (Adopted from Perdikaris et al. [77].)