Modeling unsteadiness in steady simulations with neural network generated lumped deterministic source terms

The lumped deterministic source term—neural network (LDST-NN) approach has been developed to obtain quasi-time-average solutions of cavity flows that include unsteady cavity effects in steady-state computations without the cavity. The results obtained with the LDST-NN based steady calculations are compared to the time average of fully unsteady solutions via the shear force acting on the cavity walls. Two orders of magnitude less computational time is required to obtain a quasi-time average simulation relative to time accurate simulations; a substantial savings. The estimated error, based on the calculated drag force, in these simulations is between 4% and 15% as compared to fully unsteady calculations, which is satisfactory for many design purposes. This should be compared to the 40% to 154% errors obtained by neglecting the cavity completely for these same cases. As such, the modified neural network-based LDST model is a viable tool for representing unsteady cavity effects. The LDST-NN quasi-time averaged solution was able to capture global unsteady effects correctly.; The LDSTs were found to correlate directly with observed sound pressure level trends and provide an additional means of assessing unsteadiness. The LDSTs were found to reach a maximum near the cavity/main flow interface but also extended well into the field; indicating that boundary condition representations alone would be inadequate for capturing unsteady effects.; Deterministic source terms were computed from unsteady simulations and modeled with a neural network for use in steady simulations sans cavity to capture the entire time average effect of the cavity. This was demonstrated for the entire range of Mach numbers, length-to-depth ratios and various translational velocities of the cavity wall.; The results of the study showed that modeling flow over cavities is possible with steady simulations that include source terms provided by a neural network. This method permits a considerable reduction in CPU time and is attractive for large scale simulations since it includes the effects of the unsteady phenomena without computing the unsteady flow inside the cavity.