Accelerating dynamic fluid solvers with pure data-driven deep learning models.

Numerical simulations of the Navier-Stokes equations are typically expensive and complex, which drives the need to build Reduced Order Models (ROMs) capable of predicting fluid flows. Currently, there exists frameworks where deep learning-based ROMs are coupled with numerical solvers to accelerate steady-state flow. Extending such frameworks to dynamic problems is demanding, which requires a careful handling of the coupling between the ROM and the numerical solver at each time step to prevent error accumulation. Furthermore, approaches that couple differentiable hybrid neural models with fluid solvers rely on automatic differentiation (AD), limiting integration with non-AD-supported Computational Fluid Dynamics (CFD) platforms. In this study, we present a novel non-AD-dependent approach for accelerating the dynamic fluid solver by building a framework that allows the flexibility of using the ROM prediction as an initial guess of the fluid solver at any chosen time-step. Careful consideration of the prediction length and frequency of usage of the purely data-driven ROM's output as initial guess prevents error accumulation while expediting solver convergence. The capability of the method is verified by testing it on the flow around a cylinder benchmark test case where the ROM prediction is used to accelerate fluid flow at Reynolds numbers not encountered during the training of the ROM. The methodology significantly speeds up the CFD solver while preserving the dynamical behavior of the flow. The results obtained show the potential of applying the methodology to much complex cases such as fluid solvers in fluid-structure interactions and other models that are time dependent.