Abstract

In recent years, progress in computational fluid dynamics (CFD) have made CFD-based design possible in various industrial fields that use fluids. On the other hand, however, the need for significant computation time and resources when performing fluid simulations for complex geometries has become an issue. To resolve this problem, approaches based on machine learning has been developing rapidly. In this study, we examine machine learning prediction of unsteady flow fields in complex channels. Specifically, we set an unsteady pulsating flow having one step in a channel, which causes rapid expansion or rapid contraction, as learning object, and set an unsteady pulsating flow having two steps in a channel as prediction object. In this paper, we report a result of applying convolutional neural network (CNN) + long short-term memory (LSTM) and signed distance function (SDF) to those.