GTC 2020: Learning the Hard Parts: Scaling Reacting Flow Simulations with Machine Learning
After clicking “Watch Now” you will be prompted to login or join.
Click “Watch Now” to login or join the NVIDIA Developer Program.
Learning the Hard Parts: Scaling Reacting Flow Simulations with Machine Learning
Adam Moses, Naval Research Lab
Explore a new technique for overcoming bottlenecks in computational fluid dynamics with machine learning. We'll describe using a neural chemistry solver for an order-of-magnitude speedup, opening the doors to problems that are currently impossible. Simulating chemically reacting flow — fluid flow with a chemical source term (such as a flame) — is computationally prohibitive: an at-scale combustor simulation with a practical hydrocarbon fuel can take years! Solving the chemical source term dominates this process, taking up to 90% of the total compute time. This share increases for more complex reactions, resulting in poor scalability. We'll discuss designing a neural network to model hydrogen combustion; integrating this neural chemistry solver into computational fluid dynamics code; analyzing the performance of this new approach; and discussing how these lessons could be applied to other CFD bottlenecks.