Setup for buoyancy-driven flow simulations to be studies: (top) DNS of buoyancy driven plume in stratified environment, showing the plume accelerating and then stabilizing into the neutrally buoyant layer [7]. (bottom) High-fidelity LES of thermal mixing at a pipe T-junction [8].
The enhanced understanding of the physics and the ML-based models developed during this project will help improve design, optimization and safety of advanced nuclear reactors and energy systems.
Buoyancy modulated flows are ubiquitous in natural and the built environments. Accurately predicting the behavior of these flows have implications for energy systems, nuclear reactors, wildfire plumes, thermohaline flows causing accelerated melting of icebergs, etc. Turbulence models used for simulating these phenomena will be improved using machine-learning based surrogates, that will be developed using the data generated from direct numerical simulations (DNS) of buoyancy-driven plumes and currents. DNS resolve all the relevant physics of the flow, providing invaluable data and insight into the physics of the phenomena, while being computationally expensive. Thus, the allocated compute-time on the leadership-scale supercomputers will facilitate DNS of buoyancy-driven flows across the relevant parameter range. The high-fidelity simulations will be conducted using NekRS, a high-order incompressible computational fluid dynamics solver, that can efficiently utilize the GPU accelerated supercomputers. The tera-bytes of data generated from the DNS will be used to train NeuralNetwork and ML-based turbulence models, which are expected to predict the complex physics at relatively small computational cost.
The enhanced understanding of the physics and the ML-based models developed during this project will help improve design, optimization and safety of advanced nuclear reactors and energy systems, and can be extended to improve the fidelity of models used for predicting buoyancydriven environmental flows.
