Abstract

Multiphysics simulation is the study of the interaction between multiple physical domains. An example of multiphysics simulation is Fluid-Structure Interaction (FSI). In FSI, a deformable solid structure interacts with a compressible or incompressible fluid flow. The fluid and solid domains are governed by their own principles using conservation or constitutive laws. Each domain solves its own equations separately. The domains are coupled via common boundaries by exchanging pressure and displacement of the common boundary. In this study, the coupling tool Mesh-based parallel Code Coupling Interface (MpCCI) is used to solve FSI problems. Being an integral part of various academic and industrial applications, the coupling tool is highly parameterized. The process of manually tuning the optimal parameter configuration is a tedious task given the number of parameters, types of parameters, domain expertise of the users, and the substantial time taken to solve a single simulation instance. The different choices of MpCCI parameters affect the accuracy, robustness and run-time of the co-simulations. MpCCI configurator tool is used to set parameters of MpCCI automatically employing both optimization and machine learning methods. The algorithm behind this tool consists of optimization, normalization, training and prediction. However, there are some challenges to improve the performance of the tool. In order to tune the tool, the machine learning features should be selected wisely for the training and prediction steps. Therefore, the sensitivity of the features is analysed and the most important ones are considered to be the machine learning features for the tool. Furthermore, some pre-processing methods are implemented to enhance the efficiency of training and prediction. In this work, it is investigated how effective the tool is for industrial FSI applications. The run-time obtained by the tool is compared to the run-time corresponding to the default configuration and to the actual optimal run-time as well. Additionally, it is checked whether the suggested configuration converges to accurate results.