Deep Learning for Scalable Chemical Kinetics
Alisha Sharma, Ryan F. Johnson, David A. Kessler, Adam Moses
Abstract
Chemistry is critical to many computational fluid dynamics (CFD) problems, such as propulsion system design, engine diagnostics, and atmospheric modeling. However, many real-world reacting flow problems are impractical due to the high computational cost and poor scalability of stiff chemistry integration. Scientific machine learning is gaining traction in the physical sciences as a way to overcome such bottlenecks; however, this technique is rarely used in real-world chemical kinetics codes. Artificial neural networks (ANNs) are particularly interesting due to their ability to compactly represent highly nonlinear functions. In this work, we present our ANN-based model for evolving a stiff chemical source term. We demonstrate this model with a hydrogen-oxidation reaction, integrate the model into a detailed reacting Navier Stokes CFD code, and finally discuss the potential for machine learning in reacting flow CFD.