Alireza Doostan is leading a major effort for real-time data compression for supercomputer research.

A professor in the Ann and H.J. Smead Department of Aerospace Engineering Sciences at the ÁńÁ«ĘÓƵ18, Doostan is the principal investigator on a  to change how researchers handle the massive amounts of data that result from complex physics problems like modeling turbulence and aerodynamics for air and space craft.

Compressing data is nothing new when it comes to computing, but advances in high- performance systems are now creating so much data that it becomes impossible to store for later analysis.

“Computing power has increased drastically, but moving and storing that data is becoming a bottleneck. We have to reduce the size of the data generated through large scale simulation codes,” Doostan said.

While some scientific analysis of turbulence flows can be completed faster on ever larger high-performance computing platforms, much of the information must be discarded because the scope of the data is too vast to store, making it impossible to conduct later assessments.

“There is a lot of structure and physics embedded in the data that ideally needs to be preserved to study complex flow physics or develop faster models,” Doostan said.

The goal of the grant is to both maintain accuracy of modeling data while decreasing its complexity, and critically, allowing it to be stored by compressing it in-situ, or in real-time as it is created during modeling. This is not currently possible for large-scale models, as existing technology often requires some or the entire modeling simulation be completed before compression can begin.

Joining Doostan on the project is a team of CU Boulder faculty, including Ken Jansen and John Evans, both also from Smead Aerospace, and Stephen Becker from applied math.

The team is focused on development of both traditional and deep neural models for massively parallel implementation of novel linear and non-linear dimensionality reduction techniques. It is a major undertaking, bringing together researchers with a broad range of backgrounds, including computational physics and sciences, discretization, machine learning, linear algebra, and statistics.

“This is a very interdisciplinary problem,” Doostan said. “This is not a problem one person can solve. You need a team.”

For Jansen, whose research focuses on turbulence modeling, an advance in compression could lead to significant progress across the spectrum of high-performance computing.

“This data compression research is critically important to provide access to the dynamics of our simulations,” Jansen said. “As simulations have passed petascale and are now exascale, it has become impractical to write the full solution fields to disk at a sufficient frequency and count, owing to the broad range of spatial and temporal scales of turbulence.”

The group has completed soon-to-be-published research showing strong promise for their approach. They are now working to scale up their algorithms to work at scale on supercomputing platforms like CU Boulder’s Blanca cluster as well as Department of Energy systems.

“There is still a lot to be done, but our early work has shown success and only increases the computational load by less than five percent,” Doostan said.

The three-year award runs through fall 2027. Doostan is hopeful their final product will include publicly available next-generation compression software for general use by all simulation practitioners.

Off

John Evans
Assistant Professor, Smead Aerospace
Monday, Oct. 4 | 12:00 P.M. | Zoom Webinar

Abstract: Turbulent fluid flows are characterized by a wide spectrum of spatial and temporal scales.  Unfortunately, the cost of resolving these scales with Direct Numerical Simulation (DNS) grows quickly with Reynolds number, so engineers will be unable to apply DNS to aerodynamic flows of industrial interest for many decades to come.  Alternatively, one can model all scales using Reynolds Averaged Navier-Stokes (RANS) or just the smallest scales using Large Eddy Simulation (LES).  RANS remains the turbulence modeling and simulation paradigm of choice in industry while LES continues to grow in popularity.  However, state-of-the-art RANS and LES approaches are inaccurate for many aerodynamic flows of industrial interest, especially those exhibiting flow separation or transition to turbulence.

In this talk, I will discuss our work toward arriving at improved RANS and LES approaches by leveraging advances in machine learning and the availability of high-fidelity simulation data for model training.  The key to our approach is constructing model forms with embedded invariance properties.  This enables us to train remarkably accurate, efficient, and generalizable RANS and LES models using sparse training data.  Specifically, I will provide a high-level overview of our approach as well as illustrative numerical results.  I will also highlight ongoing and future research directions including Hybrid RANS/LES modeling of separating turbulent boundary layers and in situ learning of turbulence closures from streaming simulation data.

Biography: John Evans is an Assistant Professor and the Jack Rominger Faculty Fellow in the Ann and H.J. Smead Department of Aerospace Engineering Sciences at the ÁńÁ«ĘÓƵ18.  His research interests lie at the intersection of computational mechanics, geometry, and approximation theory, with current thrusts in isogeometric analysis, immersogeometric analysis, interactive simulation, and data-driven modeling.  He has won a number of awards for his research and teaching including the 2021 Gallagher Young Investigator Award from the United States Association for Computational Mechanics and the 2021 AIAA Rocky Mountain Educator of the Year (College/University), and he is currently Associate Editor of the journal Engineering with Computers.

Off

John Evans has been named 2021 Educator of the Year by the of the American Institute of Aeronautics and Astronautics.

Evans, an assistant professor and the Jack Rominger Faculty Fellow in the Ann and H.J. Smead Department of Aerospace Engineering Sciences, is an expert in fluid dynamics and fluid-structure interaction. He teaches courses on aerodynamics and finite elements modeling at the undergraduate and graduate level.

He joined CU Boulder as a faculty member in 2013.

The AIAA RMS will recognize Evans and other 2021 award winners at a banquet at the Denver Museum of Nature and Science on August 13.

Numerous CU Boulder professors have been recognized by the AIAA RMS over the last decade. Previous recipients of the Educator of the Year Award include Hanspeter Schaub in 2020, Alireza Doostan in 2015, and Jean Koster in 2011. In 2020, Allie Anderson was also recognized as their Young Engineer of the Year.

Off