Abstract: Computational fluid dynamics (CFD) plays a key role in guiding both geometry design and physical understanding in a variety of multi-physics applications — one such application, which is the primary focus of this talk, is the design of fuel-efficient hypersonic propulsion devices (e.g., rotating detonation engines and scramjets). The required high-fidelity simulations of combustion processes in such devices encounter prohibitive computational restrictions, with the two primary challenges being (1) timescale limitations induced by detailed chemical kinetics models, and (2) issues stemming from complex geometry treatment.

The goal of this talk is to present data-driven modeling strategies to treat both challenges. For the first topic (prohibitive timescales), the central theme is high-fidelity solver acceleration — here, localized chemical kinetics models based on physics-guided data clustering strategies and neural networks are integrated into flow solvers to enable long-time simulation capability by eliminating the detailed chemistry bottleneck. For the second topic (complex geometry treatment), the central theme is development of surrogate modeling approaches based on graph neural network (GNN) autoencoders that easily interface with unstructured grid data and are not restricted to single geometric configurations after training. Emphasis is placed on showing how proper design of GNN layers can produce interpretable latent spaces in complex unsteady flows, leading to new modeling pathways compatible with complex geometry fluid applications.