Data-Driven Modeling of Compressible Reacting Flow Using Hardware-Oriented Algorithms

Abstract

High-fidelity numerical simulations of combustion processes in next-generation hypersonic propulsion devices (including, but not limited to, rotating detonation engines and scramjets) play a crucial role in enabling robust design strategies for their real-world deployment. These simulations, however, require full-geometry numerical solutions of the compressible reacting Navier-Stokes equations. Spatiotemporal resolution requirements stemming from multi-scale interactions between turbulence, shockwaves, and chemical reactions contained in these governing equations induce computationally prohibitive bottlenecks that render the required long-time resolved simulations of these propulsion devices infeasible. A particularly elusive bottleneck comes from the treatment of detailed chemical kinetics required to accurately describe the time evolution of species concentrations and flow-chemistry interactions within the combustors. The computational hurdles emerge here from the arithmetic intensity of chemical source term evaluations and immense disparities in chemical timescales for practical fuels.

The goal of this work is to provide a physics-guided data-driven modeling strategy for accelerating high-fidelity compressible reacting flow solvers via elimination of the chemistry bottleneck. Since unsteady features of interest in compressible reacting flow (e.g. detonations) are sustained by chemical reactions, the principle assumption is that local regions in the thermochemical state space can be used to classify spatially coherent regions of dynamical similarity within the reacting flowfield in physical space. Based on this assumption, the modeling approach finds these local regions using an unsupervised clustering algorithm and deploys targeted models for accelerated chemical source term evaluation within each region. The novelty comes from (a) ensuring that flowfield classifications enabled by the clustering procedure are consistent with physical expectations in complex compressible reacting flow (e.g. the clusters identify meaningful regions within detonation wave structure in rotating detonation engines), and (b) embedding physical knowledge directly into the clustering objective function. Emphasis is placed on ensuring the modeling framework can be extended to in-situ (or online) integration with flow solvers, such that the method is not tied down to single geometric configurations. Additional steps are taken to ensure that the algorithms used in the modeling approach are compatible with modern high-performance computing trends dominated by GPU-centric node architectures.