|
5-6 déc. 2024 Villeurbanne (France)
|
Conférenciers invités
Single and Multiobjective Bayesian Optimization in the presence of uncertainties
The use of numerical models is widespread nowadays to study physical phenomena, as they can be used for parameter/state estimation, or to take some decision by defining a cost function. In the modelling process, uncertainties are introduced to account for uncontrollable external effects, making the quantities of interest statistical estimates.
Due to the expensive cost of simulations, the total number of model runs is often limited. This motivates the use of surrogates which help reduce the computational cost, and can be used to choose the next point to evaluate.
In this presentation, we will go over the principles of Bayesian Optimization for single and multi objective problems, possibly constrained, and how it can be adapted when uncertainties are introduced in the modelling.
An ocean modelling framework for mitigating oceanic projections from global climate models present-day biases This paper presents an ocean-only general circulation model framework designed to (i) reduce the influence of climate models present-day biases on future ocean physical and biogeochemical projections, and (ii) assess the mechanisms driving these projected changes. The control simulation is forced by detrended air-sea fluxes from an atmospheric reanalysis, excluding climate change signals. For climate change simulations, air-sea flux anomalies diagnosed from historical and future climate model ensembles simulations, are added to these control fluxes. Heat flux anomalies are decomposed into components independent of and tied to local sea surface temperature (SST) changes, with the later modelled as an online relaxation to the control simulation’ SST. This approach results in a more realistic ocean state than in climate models, while still explicitly accounting for evaporative cooling and net longwave radiation feedback. Our results demonstrates that climate models biases can significantly compromise the reliability of projected patterns. For instance, the strong cold-tongue bias in the IPSL-CM6A-LR model leads to greater warming and chlorophyll decrease in the western equatorial Pacific, while our bias-corrected simulation shows a larger response in the eastern Pacific. Sensitivity experiments, where changes in heat, freshwater and momentum fluxes anomalies are applied separately, show that both thermodynamical (i.e., heat and freshwater-driven) and dynamical (i.e., wind-driven) processes contribute equally to this warming pattern, emphasizing the importance of Bjerknes feedback. This cost-effective method can improve oceanic projections from any climate model and offer insights into driving mechanisms. co-auteurs : S. Pang2,3, Y. Silvy4, V. Danielli1, S. Gopika1,5, K. Sadhvi1,5, C. Dutheil1, C. Rousset2, C. Ethé2, R. Person2, G. Madec2, N. Barrier1, O. Maury1, L. Dalaut1, C. Menkes6, S. Nicol7, T. Gorgues8, A. Melet9, K. Guihou9, J. Vialard2
Titre: A VENIR Résumé: A VENIR
Consideration of forcing uncertainty on the sensitivity and calibration of a catchment-scale pesticide transfer model The use of pesticides poses significant challenges to sustainable agriculture and water quality, necessitating effective risk assessment tools. The PESHMELBA model (Pesticides and Hydrology: modeling at the catchment scale) simulates water and pesticide transfer processes at the catchment scale, enabling the comparison of landscape management scenarios and their impacts on water quality. To be able to use it as a decision-making tool, it is essential to properly quantify its uncertainties, coming from various sources. While parameter uncertainty has been increasingly studied, forcing uncertainties (e.g., rainfall, pesticide application dates and quantities) are often overlooked, potentially leading to parameter calibration that cannot be extrapolated to different forcing conditions. We study the benefits of a robust approach to parameter calibration for the PESHMELBA model. To overcome the extensive computational burden intrinsic to robust calibration methods, we employ a polynomial chaos-based metamodel which approximates the response surface across parameters and emulates the uncertainty of the forcing input. We evaluate the benefits of the proposed method by comparing the robust parameter calibration with classical calibration using a set of new forcing data.
|
| Personnes connectées : 3 | Vie privée |
|
