Design of Supersonic Turbulent Combustors with US3D and the Conditional Moment Closure Model

Accurate and efficient simulation of unsteady reacting flow phenomena within liquid fuel burning hypersonic engines is an extremely challenging simulation task requiring a detailed understanding of supersonic mixing, self-ignition, liquid film atomization, detailed chemical reaction pathways, turbulent combustion interactions, shock-shock and shock-boundary layer interactions, and fluid-solid energy transfers. Though such high-fidelity models exist within commercial CFD software their computational cost is intractable for typical industrial use cases, limiting their application to solely academic use. To mitigate this limitation, reduced order models are typically adopted which employ simplifying assumptions, often decoupling interactions between physical mechanisms to a point where the desired phenomena can no longer be accurately predicted. The industry-focused engineering team at DeepFlow LLC proposes a novel solution that allows for the deployment of high-fidelity combustion modelling capabilities, at a computational cost comparable to non-reacting CFD simulations. Enabling fast and accurate simulation of hypersonic engine combustor stability regimes, including stable, unstable, unstart and blowout modes. To achieve this the Conditional Moment Closure (CMC) turbulent combustion model has been reformulated to allow for tabulated solutions of reactive scalars, computed offline, which are looked-up during the CFD simulation. ĀThough the combustion table reduces the number of numerically solved transport equations, memory demands quickly become intractable with increasing table dimensionality and resolution. However, the addition of table dimensions with high resolution is required to adequately resolve the physical dependence of reacting scalars on the various table dimensions. To address this limiting factor the DeepFlow team proposes a novel machine-learning inspired solution to reduce the memory footprint of n-dimensional combustion tables. This reduction is achieved through a two-step methodology. First an unsupervised non-linear manifold learning model is employed to reduce the dimensionality of the CMC combustion table. Second, the reduced table is then used for supervised training of various regression surrogate models. By replacing the n-dimensional table with a surrogate model, the memory demands associated with table dimensionality and resolution are relieved, and instead are solely dependent on the complexity of the surrogate model. The DeepFlow team plans to develop this technology during Phase I for production use by developing a software plugin for the commercial CFD software US3D. US3D isĀthe tool of choice to simulate complex hypersonic flow phenomenon and has been efficiently deployed on exa-scale-class clusters. By enabling US3D to perform turbulent combustion modelling, commercial users can bridge the gap between computational cost and model fidelity to simulate hypersonic engine operation both accurately and efficiently.