Fig. 1.

Training surrogate models on related data for multifidelity uncertainty quantification: (i) Simplifications are made to the physics model such as formulating it over smaller domains, ignoring some phenomena, linearizing, and stopping iterative numerical solvers early. (ii) The corresponding simplified model is simulated many times to rapidly generate large volumes of outputs that form the related training data. (iii) A surrogate model is trained on the related data from the simplified model. (iv) For realizations of the uncertain inputs such as environmental conditions, a few output samples are computed with the expensive physics model and many output samples are obtained with the affordable surrogate model. (v) The samples from physics and surrogate model are combined into unbiased multifidelity estimators of expectations and variances of the quantities of interest.