Cerfacs Enter the world of high performance ...

Realistic synthetic turbulence inflows using Deep Learning

 , ,

Required Education : Masters degree in fluid mechanics or applied mathematics
Start date : 1 October 2019
Mission duration : 3 years
Salary : 2300€ gross

Context

In the context of the Data-Science strategic program, many topics are investigated at CERFACS in relation to machine learning and computational physics. They are structured in the Helios project (High pErformance LearnIng for cOmputational phySics: https://cerfacs.fr/helios/), the cross-laboratory workgroup on machine learning with ISAE-Supaéro. One topic focuses on “AI assisted simulation”, which includes the uses of AI inside physical solvers in order to accelerate them. The issue of turbulence generation for inflows is one of the important topics of this work, as turbulent inflow generation can be done by solving the fluid equations but this is very costly in realistic simulations.

Objectives and program

The objective of this particular Ph.D. is to explore new methodologies for generating and injecting realistic synthetic turbulence in CFD simulations, based on machine learning techniques. Recent work [1] has shown the great potential for these approaches compared to traditional spectrum-based techniques, but much work is needed in this field to bring these approaches to maturity for concrete CFD applications. Data-driven approaches will be compared on 3 major typical configurations: homogeneous isotropic turbulence, wall semi-bounded and channel flows. For each, the methodologies needed to generalize to new unseen conditions after training will be explored, and sufficient training databases will be sought through Cerfacs collaborators, and/or produced locally using high-fidelity CFD.

Contacts

Corentin Lapeyre – lapeyre@cerfacs.fr

Guillaume Daviller – daviller@cerfacs.fr

Thierry Poinsot – poinsot@cerfacs.fr

References

[1] Fukami, K., Kawai, K., & Fukagata, K. (2018). A synthetic turbulent inflow generator using machine learning. arXiv preprint arXiv:1806.08903.