PhD Defense : Bastien NONY : “Reduced-order models under uncertainties for microscale atmospheric pollutant dispersion in urban areas: exploring learning algorithms for high-fidelity model emulation”
Nathalie BROUSSET | Conference room - CERFACS - Toulouse
Youtube link :https://youtu.be/MTHSD9CiF4o
In the event of an accidental release of hazardous substances in an urban area or on an industrial site, tracking pollutant concentration is particularly important for assessing public exposure to toxic doses. This is an operational but also a scientific challenge, as the interaction of the atmospheric boundary layer with the urban canopy makes the near-surface flow dynamics complex and requires high-fidelity physics modelling tools. By explicitly solving for most of the turbulence spectrum, the large-eddy simulation (LES) approach has the potential to represent the spatial and temporal variability of pollutant concentration in a complex environment. Finding a way to synthesize this large amount of information to inject into lower-fidelity operational models is particularly appealing. Still, in this accidental context, the LES approach remains subject to atmospheric and emission source uncertainties, and requires an ensemble modelling framework to represent the range of plausible dispersion scenarios. But this multi-query framework is far out of reach in a real-time context as LES simulations require very large computational resources.
In this thesis, we explore different statistical learning approaches to design a reduced-order model informed by LES to produce physically consistent concentration predictions, while substantially decreasing computational cost. This study is carried out on a two-dimensional tracer dispersion case in a turbulent atmospheric boundary-layer flow over an isolated obstacle, in which both the inflow boundary condition and source location are uncertain.
In a first step, we design a data-driven reduced-order model approach based on LES data to predict tracer concentration field statistics. We compare several dimension reduction approaches (proper orthogonal decomposition/POD versus autoencoder) to reduce the field statistics to a limited number of latent variables. We also compare several regression models (e.g. polynomial chaos, Gaussian processes) to represent the response of the latent variables to changes in the uncertain parameters. POD combined with Gaussian process regression provides fairly good predictions for a large LES training dataset (made of 450 snapshots). Near-source concentration heterogeneity upstream of the obstacle requires a large number of POD modes to be well captured. Moreover, the field dimension reduction capability of the model can be improved by replacing POD with a convolutional autoencoder.
By reducing the number of training snapshots, we observe a loss of consistency with physics principles in the reduced-order model predictions. To overcome this issue, in a second step, we design a hybrid reduced-order model approach based on a LES-informed Reynolds-averaged Navier-Stokes (RANS) tracer transport equation to integrate physical constraints in the training process. The key idea is to decouple the atmospheric uncertainties from the source location uncertainties and to replace the classical RANS turbulent closure terms with data-driven airflow models emulated from LES data. This approach requires much less LES data (only 50 snapshots) than the LES data-driven reduced-order model. We finally show that a multi-fidelity approach (combining a small number of LES snapshots with a large number of hybrid model predictions) offers an interesting avenue of research to optimize the reduced-order model performance.
Amandine MARREL – CEA – Referee
Lionel SOULHAC – INSA Lyon – Referee
Etienne MEMIN – INRIA – Referee
Laure RAYNAUD – Météo-France/CNRM – Examiner
Fabrice GAMBOA – IMFT TOULOUSE/U.P.S – Examiner
Thomas JARAVEL – CERFACS – Invited member
Didier LUCOR – LISN/Université Paris Saclay – Invited member
Mélanie ROCHOUX – CERFACS – Director