Metamodels combining CFD and measurements for plume detection and source backtracking

Status

Open

Scientific disciplines

Mathematics

Research direction

Digital Science and Technology

Affiliate site

Rueil-Malmaison

Detecting gaseous emissions from either industrial storage sites or natural environments is a major challenge when it comes to controlling harmful emissions and greenhouse gases. The challenge is not only to know precisely the dispersion of the plume formed by these emissions, and therefore the areas exposed, but also to be able to trace the source of these emissions in order to remedy them. In this context, 3D modeling enables fine resolution of local concentrations, considering the numerous parameters and highly unsteady behavior of these turbulent flows. Among the various approaches available, the LBM (Lattice Boltzmann Method) approach, combined with high-performance computing resources, is ideally suited to efficiently simulating these large-scale situations, which also present numerous sources of uncertainty. To study the impact of different parameters (wind intensity and direction, emission sources characteristics, etc.), the construction of meta-models and response surfaces makes it possible to avoid costly simulations that are difficult to implement in operational conditions.
The objective of the PhD thesis is to develop learning models or metamodels adapted to take into account both the functional nature of the quantities of interest (spatial maps of pollutants) and the differences in simulation data available, in particular through multi-fidelity techniques enabling several simulators with different calculation costs and accuracies to be taken into account. Active learning approaches will be developed to optimize the choice of simulations: simulator fidelity levels and input variable values. Finally, based on local measurements of pollutant concentrations obtained using UAVs, the inverse problem of locating the pollution source will be solved using the metamodels developed in the first part of the thesis. 

Keywords: Machine learning, multifidelity models, inverse problem, CFD, Lattice Boltzmann, pollutant dispersion

Academic supervisor    Dr Nathalie BARTOLI, ONERA, Orcid.org/0000-0002-6451-2203  
Doctoral School    ED475 MITT, Université de Toulouse
IFPEN supervisor    Dr Stéphane JAY,  Orcid.org/0000-0002-8637-0314
PhD location    IFPEN (Rueil-Malmaison) and ONERA (Toulouse), France
Duration and start date    3 years, starting in the fourth quarter 2024 (November 4)
Employer    ONERA
Academic requirements    University Master degree in applied mathematics 
Language requirements    English level C1 (CEFR), willingness to learn French   
Other requirements    Statistics, data science, optimisation, basics in Computational Fluid Dynamics

To apply, please send your cover letter and CV to the IFPEN supervisor indicated here below.

Contact
Encadrant IFPEN :
Dr Stéphane JAY