A Metamodel of the Telemac Errors

TL;DR

This paper introduces a low-cost, efficient metamodel for Telemac that approximates its results, enabling faster calibration and uncertainty analysis in hydrodynamic simulations.

Contribution

It presents a novel statistical metamodel for Telemac that reduces computational costs and aids in understanding and improving the original model's accuracy.

Findings

Metamodel significantly reduces computation time.

Metamodel accurately approximates Telemac results.

Improves calibration and uncertainty quantification.

Abstract

A Telemac study is a computationally intensive application for the real cases and in the context of quantifying or optimizing uncertainties, the running times can be too long. This paper is an example of an approximation of the Telemac results by a more abstract but significantly faster model. It shows how a metamodel can be easily built with low computational costs, and how it can help to understand and improve some global results of Telemac. I. INTRODUCTION Many sources of uncertainty lie in the real-world problems. Telemac as any model (i.e. approximation of reality) is error prone since uncertainties appear in the initial or boundary conditions, the system parameters, the modelling simplification or the numerical calculations themselves. Therefore, it is difficult to say with confidence if the design of a Telemac model has met all the requirements to be optimal. Calibration consists of tuning the model parameters so that the results are in better agreement with a set of observations. This phase is crucial before any further study can be conducted by avoiding a meaningless analysis or prediction based on false or too inaccurate results. This paper presents a statistical calibration of a Telemac 2D model (Gironde Estuary in France) with the learning of Telemac errors by a metamodel (i.e. a model of the simulation errors) to make the best use of limited observations data over a short time period. The metamodel here is a simplified version of Telemac behaving the same for all the locations where observation points are available. If the metamodel is correct, it will be able to compute as Telemac would do but with a highly reduced computational cost.