Reduced order model for chemistry in high-speed flow simulations using Fourier Neural Operators

Simulating high-speed chemically reacting flows is a computationally prohibitive task due to the burden associated with large number of species, stiff chemical reactions, and mesh resolution requirements. Using reduced order models (ROMs) for chemically reacting flows promise to alleviate some of this computational burden. A possible ROM could be disruptive if a reduction of chemical state and mesh size can be achieved. In this work, we target the minor species which occupy only a small percentage of the actual simulation domain but are critical for accurate prediction. We utilize a Fourier Neural Operator (FNO) based input-output model as a possible ROM technology. We use the FNO model to predict the minor chemical species present in a Direct Numerical Simulation (DNS) using velocity fields, key thermodynamics, and the major chemical species as inputs. We show that our FNO model is able to learn the profile of a steady state 1D flame front for a range of pressures, temperatures, and equivalence ratios. We also applied FNO-based predictions to unsteady three-dimensional flows, including chemically reacting Jet-in-Crossflow (JICF), to demonstrate the method's applicability to complex configurations.