Online Machine Learning for Large-Scale Turbulent Simulations

PI Kenneth Jansen, University of Colorado Boulder
Co-PI Jed Brown, University of Colorado Boulder
John Evans, University of Colorado Boulder
Alireza Doostan, University of Colorado Boulder
Stephen Becker, University of Colorado Boulder
Jansen INCITE Graphic

Direct numerical simulation of separated flow over a bump provides valuable training data on the effects of pressure gradients and wall curvature. Three stream wise slices at x/L={0.1, 0.2, 0.3} and one span wise slice show development of the separation and reattachment regions of the flow. Image: Kenneth E. Jansen, University of Colorado Boulder

Project Summary

This project will advance the current state of the art for online data analytics and machine learning applied to large-scale computational fluid dynamics (CFD) simulations to develop enhanced turbulence models for flows of interest to the aerospace, automotive, and renewable energy industries.

Project Description

This project is motivated by the fact that the vertical tail of a commercial airplane is a significant contributor to the overall drag and fuel cost during cruise, and the sizing of this component is dictated by engine-out operations which require large surfaces to produce the necessary restoring force. A more effective system that uses flow control can utilize smaller surfaces, resulting in reduced drag and fuel cost during standard operations ($0.3B per year for large commercial airlines). 

Building upon current and previous work leveraging DOE supercomputers, the team will use this INCITE project to advance the current state of the art for online data analytics and machine learning applied to large-scale computational fluid dynamics (CFD) simulations. They plan to develop more predictive lower fidelity (and thus less computationally expensive) turbulence models for flows of interest to the aerospace, automotive and renewable energy industries. Through the integration of a new flow solver designed for GPUs with distributed and online data analytics and training algorithms, the team’s research will greatly enhance the confidence in lower fidelity models and enable engineers to obtain more accurate solutions to complex flows outside the reach of today’s modeling capabilities. 

The team’s work aims to extend neural net sub-grid stress (SGS) models for large eddy simulation (LES) beyond canonical turbulent flows. By continuing their prior flat plate direct numerical simulation (DNS) within a new GPU-based solver coupled with online learning of wall-bounded flows with increasing complexity and scale, the researchers can provide training data for SGS closures that is currently unavailable to the community. Using the DNS of a boundary layer over flat plate, they will develop an SGS neural net model capable of accurately predicting flows of increasing complexity. Finally, to evaluate the trained SGS model on a previously unseen flow, they will perform LES of the turbulent boundary layer over an airfoil with flow separation and a second LES of a vertical tail/rudder assembly. This is a particularly relevant flow case for the aerospace and renewable energy industries, therefore making a predictive closure extremely valuable.

Allocations