Deep Spatio-Temporal Learning of Complex Naval Fluid Systems

The design of naval vehicles and missions requires a detailed analysis of fluid interactions through computational fluid dynamics (CFD). These simulations are often extremely slow and cumbersome due to the complexities of the governing fluid physics, the required spatial grid to resolve key effects, and the required temporal resolution for computational stability. However, recent advances in deep learning suggest a potential for a dramatic shift in the CFD paradigm. These methods suggest that the solution to complex stochastic partial differential equations (SPDEs), including the Navier-Stokes equation, can be learned. Furthermore, once trained, the learned solution can have a speedup on the scale of 3 to 4 orders of magnitude over direct numerical solvers. In this work we attempt to substantiate these bold claims specifically in the untested realm of CFD, and discuss the relevant trade-offs between learning methods.