Prediction of Hydrogen Mixing and Velocity Fields Behind a Strut Injector with Multi-Lobe Nozzles at scramjet engine using POD+LSTM technique

This work proposes a hybrid data-driven reduced-order modeling (ROM) approach that combines Proper Orthogonal Decomposition (POD) with Long Short-Term Memory (LSTM) networks to efficiently predict unsteady hydrogen mixing and supersonic flow fields downstream of a strut injector. The study considers three annular nozzle configurations—2-lobe, 3-lobe, and 4-lobe—under high-speed, non-reacting hydrogen injection conditions. The full-order model (FOM) data used for training and validation were generated via Unsteady Reynolds-averaged Navier–Stokes (URANS) simulations employing the turbulence model, capturing the transient flow features and scalar transport in detail. POD was employed to obtain the dominant spatial modes from the computational results, while LSTM networks were trained on the temporal evolution of the modal coefficients to forecast the flow and scalar fields. The ROM performance was evaluated under various training-to-testing ratios (70%, 80%, and 90%), and the results were benchmarked against full-order contours of hydrogen mass and Mach number on a representative downstream plane. The proposed POD+LSTM framework demonstrated accurate predictions when trained with at least 80% of the dataset, with near-exact reconstruction achieved at 90% training. Contour comparisons showed that both the velocity and scalar fields were well captured in terms of jet penetration, shock structures, and mixing layer development, particularly for the more complex 3-lobe and 4-lobe nozzles. The results demonstrate the potential of POD–LSTM as an efficient reduced-order tool for rapid prediction of unsteady hydrogen mixing and flow structures in high-speed fuel-injection systems, with substantially reduced computational requirements relative to repeated full-order CFD simulations.