A MULTI-FIDELITY REDUCED-ORDER MODEL APPLIED TO RESER- VOIR ENGINEERING

Resumo

Although the rapid evolution of computational power, some applications (e.g., optimization, uncertainty
propagation), combined with the increase of simulation complexity, may still require computational times that are
unsuitable for practical purposes. Therefore, in the present work, a multi-fidelity surrogate strategy for expensive
reservoir engineering numerical models is presented. The main idea of the approach is to combine information
from few high-fidelity samples (of long computational duration) and accurate simulations with faster evaluations
from a low-fidelity model with lower accuracy. This strategy provides enough information for creating a surrogate

model, however with a lower computational cost than what would be obtained by evaluating just high-fidelity sam-
ples (for the same level of quality). Furthermore, conversely to the majority of multi-fidelity models, the present

strategy includes a dimensionality reduction technique to deal with multi-dimensional outputs. The presented ap-
proach is employed in an example of a reservoir engineering problem for predicting both the net present value and

the saturation of oil in the reservoir. The results obtained are suitable when evaluating both time and accuracy,
indicating that the presented approach is promising for practical applications. Moreover, it also suggests that the
same framework may be used in other computationally demanding applications.