Abstract: Numerical investigation of high-pressure fluid flows has traditionally relied on solving the conservation equations of mass, momentum, total energy and species for a multi-component mixture, coupled to a generalized equation of state (EOS) valid for the entire thermodynamic régime. These computations, however, are often too expensive and time-consuming, mainly due to the highly nonlinear nature of the EOS, the coupling between multiple physio-chemical phenomena, and the grid-resolution requirements. The cost is further exacerbated in the case of design-oriented studies, for which a large number of parametric variations is needed to survey the full design space.
This talk investigates these challenges and presents a systematic approach to achieve simulation results in a reasonable turnaround time by leveraging recent advances in machine learning techniques. The efficiency of evaluating real-fluid properties, a computationally expensive procedure, is dramatically enhanced using deep feedforward neural networks. In addition, a new design paradigm is proposed based on fast-running emulators that approximate the multivariate input/output behavior of complex systems, hence allowing for rapid exploration of the design space. The emulation framework assembles a set of techniques, including design of experiments (DoE), high-fidelity large eddy simulations (LES), reduced-order modeling and deep learning. As specific examples from work carried out at Argonne and Georgia Tech, the flow dynamics in a single-hole diesel injector and swirl rocket injector are discussed.