Abstract

Today's most challenging problems in numerous technical disciplines involve multiple complex, interacting physical systems, often incorporating different energy exchange mechanisms involving several states of matter. One such domain is fluid-structure interaction (FSI) where fluid and solid structure interaction takes place. Even with the computational advances we have seen in the last couple of decades, large-scale high-fidelity FSI simulations are still cost prohibitive for industrial use. Solving for the phenomena of interest in a given simulation domain requires a certain mesh density. Without the necessary a-priori simulation knowledge, the user is unable to decide the exact areas of interest to refine before the simulation. Solving the problem on a uniformly dense mesh becomes expensive and time-consuming. Adaptive mesh refinement (AMR) technique enables the drastic reduction in computational costs and memory requirements for large-scale multi-physics simulations by refining and coarsening the mesh on the fly where needed during the simulation. Adaptive mesh refinement requires users to manually select the refinement or coarsening thresholds which in turn requires intimate knowledge about the problem and a-priori solution. This limits the usability of adaptive mesh refinement for the broader simulation community. FSI problems involving deforming or oscillating structure inherently requires mesh motion capabilities. Currently, in the open-source software community, a generalized, combined capability of performing adaptive mesh refinement and mesh motion does not exist. To provide a solution to this problem, Illinois Rocstar LLC is developing a generalized mesh motion and machine learning-based smart adaptive mesh refinement framework for FSI simulations with IR's in-house multi-physics simulation suite "Rocstar Multiphysics" and in-house meshing software NEMoSys. NEMoSys smart adaptive mesh refinement and mesh motion capabilities are linked against Rocstar Multiphysics FSI solver to demonstrate the proposed capability. This capability will allow users to perform FSI simulations while leaving adaptive mesh refinement threshold selection and application to the pre-trained machine learning model. The state-of-the-art machine learning network trained on several high-fidelity CFD and FEA problems can correctly identify refinement and coarsening regions in a domain based on the current time-step solution field. This capability is tested on a real-world 3D FSI problem of a wind turbine blade undergoing significant wind loading and deformation/oscillation. Further, we also make assessments on the stability of this combined mesh motion and adaptive mesh refinement capability.