Multifidelity Surrogate Models for Efficient Uncertainty Propagation Analysis in Salars Systems

Salars are complex hydrogeological systems where the high-density contrasts require advanced numerical models to simulate groundwater flow and brine transport. Applying those models over large spatial and temporal scales is important to understand the various subsurface processes in salars, but the associated computational cost hinders an analysis based on repetitive numerical simulations. Single fidelity surrogate modeling is a common approach to alleviate computational burden with computationally expensive physics-based models of high-fidelity. However, due to the complexity in salars modeling it might not be affordable to run high-fidelity simulations many times until we build a surrogate model of acceptable accuracy. Here, we investigate if multifidelity surrogate methods, that exploit information from inexpensive lower fidelity models, can show promise for computationally demanding tasks for salars systems. Additive, multiplicative and co-Kriging multifidelity surrogates are developed based on the combination of training data from low fidelity sharp interface models and a higher fidelity variable-density flow and solute transport model. Their performance is compared against a single fidelity Kriging surrogate model, and they are all employed to conduct a Monte-Carlo-based uncertainty propagation analysis where recharge, hydraulic conductivity and density differences between freshwater and brine are considered uncertain model inputs. Results showed that multifidelity methods are a promising alternative for time-intensive numerical models of salars under limited high-fidelity samples. In addition, sharp interface models, despite commonly used in coastal aquifer problems, can also be applied in salars modeling as cheap lower fidelity models for interface calculations via a multifidelity framework. The Monte-Carlo outputs based on the surrogate models, resulted in estimated probability density functions characterized by long tails, thus, highlighting the need to reduce parametric uncertainty in real world models of salars.