Model emulation and moment-independent sensitivity analysis: An application to environmental modelling

  • Authors:

  • E. Borgonovo;W. Castaings;S. Tarantola

  • Affiliations:

  • DEC and ELEUSI, Bocconi University, Via Roentgen 1, 20136 Milano, Italy;Lab. EDYTEM (CNRS, Université de Savoie), Campus scientique, F-73376 Le Bourget du Lac, France and Université de Toulouse, INPT, UPS, IMFT, Institut de Mecanique des Fluides de Toulouse, ...;Institute for the Protection and Security of the Citizen, Joint Research Centre of the European Commission, TP 361, Via E. Fermi, 2749 - 21027 Ispra, Italy

  • Venue:
  • Environmental Modelling & Software
  • Year:
  • 2012

Quantified Score

Hi-index 0.00

Visualization

Abstract

Moment-independent sensitivity methods are attracting increasing attention among practitioners, since they provide a thorough way of investigating the sensitivity of model output under uncertainty. However, their estimation is challenging, especially in the presence of computationally intensive models. We argue that replacement of the original model by a metamodel can contribute in lowering the computation burden. A numerical estimation procedure is set forth. The procedure is first tested on analytical cases with increased structural complexity. We utilize the emulator proposed in Ratto and Pagano (2010). Results show that the emulator allows an accurate estimation of density-based sensitivity measures, when the main structural features of the original model are captured. However, performance deteriorates for a model with interactions of order higher than 2. For this test case, also a kriging emulator is investigated, but no gain in performance is registered. However, an accurate estimation is obtained by applying a logarithmic transformation of the model output for both the kriging and Ratto and Pagano (2010) emulators. These findings are then applied to the investigation of a benchmark environmental case study, the LevelE model. Results show that use of the metamodel allows an efficient estimation of moment-independent sensitivity measures while leading to a notable reduction in computational burden.