Supercomputing

Machine Learning to Detect Convergence for Space Launch System CFD Simulations

Image from animation depicting machine learning model working on SLS CFD
Animation showing the classification of computational fluid dynamics iterative history. The figure at the top left depicts surface pressure overlaid on NASA’s Space Launch System and the surrounding flow field solution. Purple represents subsonic flow, green represents supersonic flow below free stream velocity, and orange represents supersonic flow above free stream velocity. The bottom left figure displays the value of an aerodynamic coefficient at each iteration of the simulation. The flow chart at right depicts the model used to classify the iterative history of the aerodynamic coefficient. Joshua Diaz, Derek Dalle, NASA/Ames

As a pivotal part of NASA’s Artemis program, the Space Launch System (SLS) has become one of the most extensively simulated launch vehicles in the history of space exploration. To accomplish this feat, a team of aerodynamics engineers conducted thousands of computational fluid dynamics (CFD) simulations for ascent and booster separation of the SLS, harnessing the high-performance computing power offered by the NASA Advanced Supercomputing (NAS) facility at Ames Research Center. A critical and challenging component of conducting CFD simulations is determining when the simulation has sufficiently converged on a solution. The team utilized recently constructed aerodynamic databases and deep neural networks that were trained to classify CFD simulations in the same manner as professional human engineers. By pairing state-of-the-art artificial intelligence models with existing convergence criteria methodologies, their tool develops a standard for precision and allows more rapid development of aerodynamic databases.

Quick Facts

Automating CFD convergence detection with deep neural networks marks a transformative leap toward precision and efficiency in NASA’s Artemis Program, and sets a foundation for larger implementation of artificial intelligence in the day-to-day work of a NASA engineer.

Joshua Diaz,
NASA Ames Research Center
  • Each neural network is trained on thousands of CFD simulations that were run on the Pleiades, Electra, and Aitken supercomputers at the NAS facility over the duration of the SLS program, using the NASA-developed CFD code FUN3D.
  • Developing an automated approach to judging CFD convergence allows for more rapid development of aerodynamic databases for NASA’s Artemis program and standardizes the process across a team of multiple analysts.
  • They found that results extracted from CFD simulations run by a seasoned engineer and CFD simulations run by the artificial intelligence models are nearly identical.
  • Future work is focused on generalizing the neural networks to work on a broad range of aerodynamic problems, as well as with multiple CFD solvers such as OVERFLOW and Cart3D.

Researchers

  • Joshua Diaz, NASA Ames Research Center
  • Derek Dalle, NASA Ames Research Center

More Information