Accelerating mixing controlled turbulent combustion simulations with hybrid Navier–Stokes/ANN scalar-solvers

Despite advances in computing power, a major limitation in the simulation of turbulent flame stems from the need to track all chemical species involved in the thin reaction zones throughout the flow field. This paper investigates how Reduced Order Models (ROMs), combining data-driven analysis and neural network training, can significantly reduce computational cost. Specifically, neural networks are employed to assist in solving ϕ(x̲,t), a thermochemical scalar representing species mass fractions, energy, or temperature. The evolution of ϕ(x̲,t) over no time steps is used as input to a ROM framework, in which dimensionality reduction is achieved using Proper Orthogonal Decomposition (POD), while temporal dynamics are modeled using a Long Short-Term Memory (LSTM) network, with ANN trained for each of the retained POD modes. The scalar field for the nrom subsequent time steps is then predicted by the network, bypassing the need to solve the transport equation for these iterations. In this work the pair of values (no=10,nrom=1) and (no=20,nrom=5) are implemented. This approach is first validated on a non-reactive Large Eddy Simulation (LES) of a cavity flow, where air and H2 are injected separately and mix downstream. The methodology is then extended to a reactive Unsteady Reynolds-Averaged Navier–Stokes (URANS) simulation of a non-premixed H2-air flame stabilized downstream of the same cavity geometry, assuming infinitely fast chemistry. When skipping CFD iterations, the network can also predict the flow evolution over a time step that is ten times larger than the standard CFD time step. This leads to a reduction in computational cost to reach a given physical time. Results demonstrate that the ROM is capable of accurately predicting the unsteady dynamics of the turbulent system across testing sequences unseen during training. The approach yields a CPU time saving of the order of 27%.