A Computationally Efficient Method for Estimating Multi‐Model Process Sensitivity Index. Issue 10 (26th October 2022)

• Record Type: Journal Article
• Authors: Dai, Heng; Zhang, Fangqiang; Ye, Ming; Guadagnini, Alberto; Liu, Qi; Hu, Bill; Yuan, Songhu
• Is Part Of: Water resources research. Volume 58:Issue 10(2022)
• Page Start: n/a
• Page End: n/a
• Publication Date: 2022-10-26
• Subjects: multi‐model process sensitivity index -- global sensitivity analysis -- quasi‐MC method -- binning method -- process model uncertainty -- parametric uncertainty; Hydrology -- Periodicals; 333.91
• DOI: 10.1029/2022WR033263
• Languages: English
• ISSNs: 0043-1397

Abstract

Abstract: Identification of important processes of a hydrologic system is critical for improving process‐based hydrologic modeling. To identify important processes while jointly considering parametric and model uncertainty, Dai et al. (2017), https://doi.org/10.1002/2016WR019715, developed a multi‐model process sensitivity index. Numerical evaluation of the index using a brute force Monte Carlo (MC) simulation is computationally expensive, because it requires a nested structure of parameter sampling and the number of model simulations is on the order of N 2 ${N}^{2}$ ( N being the number of parameter samples). To reduce computational cost, we develop a new method (here denoted as quasi‐MC for brevity) that uses triple sets of parameter samples (generated using quasi‐MC sequence) to remove the nested structure of parameter sampling in a theoretically rigorous way. The quasi‐MC method reduces the number of model simulations from the order of N 2 ${N}^{2}$ to 2 N . The performance of the method is assessed against the brute force MC approach and the recent binning method developed by Dai et al. (2017), https://doi.org/10.1002/2016WR019715, through two synthetic cases of groundwater flow and solute transport modeling. Due to its rigorous theoretical foundation, the quasi‐MC method overcomes the limitations imposed by the inherently empirical nature of the binning method. We find that the quasi‐MC method outperforms both the brute force MC and the binning method in terms of computational requirements and theoretical aspects, thus strengthening its potential for the assessment of process sensitivity indices subject to various sources of uncertainty. Key Points: A new quasi‐Monte Carlo (MC) method is developed for more efficiently estimating the multi‐model process sensitivity index The new method reduces the number of MC model simulations for estimating process sensitivity index from the order of N 2 to 2 N This mathematically rigorous new method also outperforms the empirical binning method in terms of accuracy and convergence