Shivam Barwey is a Ph.D. candidate in the Department of Aerospace Engineering. He is a member of the Advanced Propulsion Concepts Lab advised by Prof. Venkat Raman. His research interest lies in accelerating high-fidelity multi-physics flow simulations with AI-based methods. Since new high-performance computers (HPCs) are primed for GPU-optimized machine learning (ML) applications, using ML-derived tools to replace existing expensive algorithms in traditional high-fidelity solvers offers an elegant pathway for GPU offloading. As a result, due to the extremely high peak performances achievable on GPU-dominant HPCs, long-time detailed simulations of full-geometry devices/domains of interest become more realizable.
To this end, Shivam’s recent work involves the development of a GPU-based library to accelerate compressible reacting flow solvers, where one primary task is to reduce the computational burden associated with chemical source term evaluations. The main workhorse is the artificial neural network (ANN), which, if properly constrained by the underlying physical mechanisms at play, allows for both highly efficient and accurate GPU-based execution. The library routines have been integrated into an in-house multi-physics solver to accelerate, using several thousand GPUs, full-geometry rotating detonation engine (RDE) simulations on the ORNL Summit supercomputer. Shivam’s current work involves extending these methods to enable fast execution in different multi-physics, multi-scale settings, such as pollutant dispersion in urban environments.