A Data-centric Approach to Turbulence Simulation

University Of Colorado At Boulder, Boulder CO

Investigators

Abstract

Turbulence or turbulent flow is commonly observed in many everyday phenomena, such as airflow over a vehicle, billowing clouds, rising smoke or fast-flowing waters. Understanding and accurately modeling turbulent flow is important for design in applications as diverse as transportation (planes, cars, etc.), energy generation (wind turbines or more traditional power plants), and manufacturing. Turbulence evolves through interaction of a broad range of length and time scales, posing significant challenges to engineers who model turbulent fluid flow using computer simulations. High fidelity computational models, which resolve turbulent motions, can require several months of time on the largest supercomputer to simulate even a relatively simple flow. By contrast, low fidelity models that attempt to predict only the net effects of the turbulence can give results on complicated flows in seconds but the prediction has low accuracy for many important flows.  To effectively use simulations of turbulence for design applications, engineers typically require flow solutions for thousands of problem-defining parameter combinations. The need for a large number of cases typically necessitates the use of lower fidelity models to keep cost and time feasible. This proposal involves using the more detailed information from a select, small number of higher fidelity simulations together with a full set of inexpensive low fidelity simulations to collectively achieve a substantially more accurate prediction for the thousands of parameter variations needed. The scale-resolving simulations proposed lend themselves to animations of fluid structures. While a full understanding is beyond students at the K-12 levels, simulations that show macro- and small-scale fluid structures are expected to be highly motivating to science, technology, engineering, and mathematics (STEM) areas. Therefore, visualization results from this research are being packaged for direct outreach activities in coordination with the Broadening Opportunity through Leadership and Diversity (BOLD) Center at the University of Colorado Boulder. This research project involves three main objectives: 1) using data-driven machine learning from scale-resolving simulations to improve scale-modeling closures, 2) developing multi-fidelity models that fuse data from O(1000) low-cost simulations and O(20) higher-cost simulations to achieve an accuracy approaching that of the higher-cost simulations for all O(1000) parameter combinations at substantially reduced cost, and 3) combining these two approaches. These developments are being demonstrated on flows that are currently known to be beyond the predictive capacity of the low-cost models. The focus is on adverse pressure gradient flows, including those that lead to flow separation. The simulations performed in this project will advance the state-of-the-art in adverse pressure gradient flow simulation. Specifically, simulations are being executed at momentum Reynolds numbers of 6,000 for direct numerical simulation, 30,000 for wall-resolved large eddy simulation, and 150,000 for wall-modeled large eddy simulation. The data from all simulations is being be made available to the science and engineering community to enable further analysis and fundamental insight.