Skip to main content
NASA Advanced Supercomputing (NAS) Division World-class modeling, simulation, and supercomputing in support of the NASA mission

Bayesian Machine Learning for Data-Enhanced CFD

AMS Seminar Series

Speaker: Paola Cinnella, Sorbonne University

Presentation

Seminar Slide Deck (PDF-7.4MB)

Abstract

In the last decade or less, applications of data science, and more specifically of so-called machine learning (ML), in physical sciences have been growing exponentially. ML-assisted Computational Fluid Dynamics (CFD) models with magnified predictive capabilities and computational efficiency can play a game changing role in Aerospace and Energy industries. However, several criticalities must be addressed for ML-enhanced CFD scaling-up to large problems of Engineering interest, including the difficulty of dealing with sparse/noisy data, the lack of generalization to out-of-sample environments, and the need for quantifying prediction uncertainties. In my talk, I will adopt the point of view of Bayesian machine learning. Such a framework is well-suited for scarce/noisy data, it enables simultaneous model selection and inference, and it provides estimates of uncertainty in the predicted model outputs, but it can be very costly to train. Efficient Bayesian algorithms are then introduced and implemented in the context of data-driven learning of turbulent closures. Their potential is demonstrated for test cases from the recent NASA Turbulence modelling Symposium, as well as for an illustrative turbomachinery flow configuration.

Biography

Photo of Paola Cinnella
Paola Cinnella is a full professor in Fluid Mechanics at Sorbonne University in Paris and member of the Jean Le Rond D’Alembert Institute of Theoretical and Applied Mechanics. Graduated in Mechanical Engineering from Politecnico di Bari in Italy in 1995, she received his Ph.D. in Fluid Mechanics from the Ecole Nationale Supérieure d’Arts et Métiers in Paris in 1999. Her research field is in Computational Fluid Dynamics and related topics, including optimization, uncertainty quantification, and data-driven modelling, with application to compressible and turbulent flows in Aerospace and Energy applications. She is associate editor for Computers & Fluids, editorial board member for Flow, Turbulence and Combustion and Scientific Reports, scientific Secretary of the International Conference in Computational Fluid Dynamics (ICCFD) and coordinator of the Scientific Interest group “Machine Learning in Fluid Dynamics” of ERCOFTAC (European Research Community in Flow, Turbulence and Combustion).