Turbulent Flows with Multi-Physics

Wall-bounded Turbulent Flows

Turbulent Poiseuille flow at Reτ = 5200Picture
Turbulent Poiseuille flow at Reτ = 5200
Turbulent Couette flow at Reτ = 500
Turbulent Couette flow at Reτ = 500
Vortical structures of turbulent Poiseuille flow at Reτ = 180
Transition from laminar flow to turbulent flow at Reτ = 180
Reynolds number dependency; Reτ = 180 (top), 550 (middle), 1000 (bottom)

Turbulent Convection

Rayleigh-Bènard Convection (Ra = 6e6, Pr = 1)

Reactive Turbulent Flows

Ammonia/Hydrogen/Nitrogen Combustion
(Video from APS-DFD Gallery of Fluid Motions, 2021)

Hypersonic Flows

Simulation of Apollo 13 re-entry
​(Courtesy - M. Gallis, Sandia National Labs)
Kelvin-Helmholtz instability in hypersonic flows
(Left) Ma = 0.3 (Right) Ma = 3.0

Multi-Phase Turbulent Flows

To be updated.

Magnetohydrodynamics (MHD) Turbulent Flows

To be updated.

Data-driven Science and Engineering

Operator Regression with Physics-Informed Neural Networks

Pressure from regressed EOS from DSMC data using neural network with regularization
Pressure from regressed EOS from DSMC data using neural network with regularization. Perfect gas for argon (Top left). Learned EOS using data from one (Top right), two (Bottom left), and four (Bottom right) Riemann problems.

Data Reconciliation with Convolutional Neural Networks

High-Performance Computing

Summit HPC system, Oakridge National Lab (Image from olcf.ornl.gov)Summit HPC system, Oakridge National Lab (Image from olcf.ornl.gov)



​Turbulent flows are a fundamentally multi-scale phenomenon. The spectrum of length-scale becomes wider with Reynolds number. Therefore, the first principle-based simulation requires intensive computing capability. We study algorithms and their implementations to maximize the computing efficiencies in modern high-performance computing systems.
Strong scaling benchmark result of our in-house DNS software
Strong scaling benchmark result of our in-house DNS software. (Numbers next to the supercomputer name are the degree of freedom for benchmark cases.)
​The benchmark result of our in-house simulation software shows great performance from a hundred processors to a million processors. Now, we are developing a new simulation software that will be efficient in state-of-the-art computing systems with heterogeneous architectures.