Online Machine Learning for Large-Scale Turbulent Simulations

PI Kenneth Jansen, University of Colorado Boulder
Co-PI Kenneth Jansen, University of Colorado Boulder
John Evans, University of Colorado Boulder Jed Brown, University of Colorado Boulder
Jed Brown, University of Colorado Boulder
Stephen Becker, University of Colorado Boulder
Alireza Doostan, University of Colorado Boulder
Sam Partee, Hewlett Packard Enterprise
Jansen Graphic

Isosurface of instantaneous Q criterion colored by speed over a vertical tail at Re = 3.5 105, with a rudder deflection angle of 30◦and a single unsteady jet active (fifth from the top). This DDES simulation on a 5 billion element mesh, run on 64k KNL cores on Theta, shows our ability to refine the grid to capture the unsteady structures resulting from the separation near the rudder and from the interaction between the unsteady jet and the crossflow.

Project Summary

This INCITE project seeks to advance the current state of the art for online data analytics and machine learning (ML) applied to large-scale computational fluid dynamics simulations, as well as 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.

Project Description

Building on work being performed as part of Argonne’s Aurora Early Science Program (ESP), the project seeks to advance the current state of the art for online data analytics and machine learning (ML) applied to large-scale computational fluid dynamics simulations, as well as 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 mature flow solver already scaled to more than 3 million processes with distributed and online data analytics and training algorithms, this work 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.

Using tools developed under the ESP allocation to extend beyond canonical turbulent flows neural net sub-grid stress (SGS) models for large eddy simulation, (LES) the researchers will perform two direct numerical simulations coupled with online learning of wall-bounded flows with increasing complexity and scale so as to provide training data for SGS closures. Hence the team will develop a neural net SGS model capable of accurately predicting flows of increasing complexity and that surpasses current state-of-the-art closures. LES on the same flows computed by DNS will validate the accuracy of the newly trained SGS model. Finally, to evaluate the trained SGS model on a previously unseen flow, the team will perform LES of the turbulent boundary layer over an airfoil with flow separation—a particularly relevant flow case for the aerospace and renewable energy industries, therefore making a predictive closure extremely valuable.

Allocations