HPC Impact Showcase:
HPC and AI - Accelerating Design of Clean High-Efficiency Engines
HPC Impact Showcase
TimeTuesday, 19 November 20194:15pm - 5pm
DescriptionSupercomputers have long been utilized in scientific fields in “capability” fashion, i.e., for performing scale-resolved simulations in a short time. However, very few real-world applications have been demonstrated using a supercomputer in “capacity” computing fashion, i.e., by running many high-fidelity simulations simultaneously. The presentation will demonstrate how an IBM Blue Gene/Q supercomputer at Argonne National Laboratory was utilized in “capacity” fashion to optimize a heavy-duty engine running gasoline for improved efficiency. Combustion engines are extremely challenging to accurately simulate due to the disparate length and time scales, combined with a multitude of physical sub-processes (injection, evaporation, etc.), intertwined with complicated fuel and emission chemistry. A high-fidelity simulation approach was first developed and incorporated into the CONVERGE CFD code, and validated against experiments. A first-of-its-kind study was performed to guide engine design by exploring a large design space and operating range. Thousands of high-fidelity engine design combinations that would ordinarily require months on a typical cluster were simulated within days on Mira. The accelerated simulation time allowed the evaluation of an unprecedented number of variations within a short time span. Machine Learning was used to develop a surrogate model for the simulations using a “superlearner” approach. This further offered the potential for reducing design time for subsequent iterations. Finally, significant fuel efficiency improvement and lower emissions were demonstrated using the real-world hardware recommended by the simulation. The ability to perform many high-fidelity calculations of a complex device on a supercomputer reduces time-to-science and opens a new frontier in the automotive industry.