Figure 4.

A pedagogical example: the NARGP algorithm can capture complex nonlinear, non-functional and space-dependent cross-correlations by inferring the nonlinear latent manifold $\mathcal{G}$ that governs the functional relation between the inputs x and the outputs of the low- and high-fidelity models f_l(x) and f_h(x), respectively. (a) The low-fidelity model is projected onto the nonlinear latent manifold $\mathcal{G}$ that is inferred using the deep non-parametric representation of equation (2.11). (b) The high-fidelity function f_h(x) is recovered by a smooth mapping from the $\mathcal{G}$ manifold to the high-fidelity data. (Online version in colour.)