AI/ML in Turbulence Modeling (Recent Advancements)
Personal Favorite: Physics-Informed ML to invoke physical constraints.
Data-driven turbulence modeling has been a very active area of research for quite some time. Admittedly, one of my first ‘loves’ when getting into data science. AI/ML has offered a variety of advancements to help our turbulence modeling capabilities nudge forward.
Don’t forget the importance of the problem (so anything helps):
“Turbulence is also the primary cause of viscous drag with 30% of the energy consumption worldwide being spent on overcoming drag in transportation” [1]
Without re-inventing the wheel and starting from scratch, I wanted to highlight some papers/innovations in this area since Jan. 1 2024. I am sharing my thoughts here as I ‘study up’ on recent advancements in this field before giving my lecture at the Johns Hopkins Applied Physics Lab (which includes this topic).
Thanks for being here - let’s dive in into ways it’s being used, papers/PDFs, and any code provided my our references. I will start with those papers cited most heavily in this time period (Jan. 1 ‘24 to present).
Data assimilation & ROMs under physics constraints
One key hurdle industry is still figuring out is how to make ML model predictions obey laws of science or known behaviors for specific problems (fluid dynamics). Conservation of energy, empirical relationships we develop for our fleets over time, basic laws of physics, among others. It will be quite enabling once there are more agnostic tools we can deploy, into a variety of ML architectures folks may be using, to ensure our model predictions have sanity checks in place for physics.
In this paper, turbulence models augmented by PINNs reconstruct mean fields from sparse sensors using underdetermined RANS equations, frequently beating simple baseline turbulence models in the same setup. Link (PDF below, this is also image source).
