Machine-learning-based defiltering for particle tracking in curved turbulent channel flow

We assess the robustness of our defiltering method for particle tracking simulations in coarse-grid fields using turbulent channel flow with streamline curvature. In practical applications, numerical simulations utilize affordably coarse grid resolutions for flow fields to reduce computational costs. However, when discrete particles are introduced into the computational domain, coarse-grid flow fields lead to inaccurate particle behavior. While several defiltering methods have been proposed to reconstruct accurate fluid velocity in coarse-grid fields, their application to practical engineering remains limited. One reason is that the validation of these methods has been limited to simple configurations such as plane channel flows and homogeneous isotropic turbulence. In this study, the machine-learning-based defiltering method proposed in our previous work [T. Oura and K. Fukagata, Phys. Fluids 36, 113366 (2024)] is applied to flows with streamline curvature because most practical targets involve curved geometries. The machine-learning model trained in a minimal plane channel is applied to the channels with three different curvatures to perform particle tracking, and several particle statistics, including the mean velocity, velocity fluctuations, migration toward the walls, and dispersion, are investigated. Notably, we reveal that a smaller stencil for the model inputs shows more favorable results than a larger one. Based on the analysis of two-point correlations for fluid velocity fluctuations, this can be attributed to the fact that local features in curved channels are no longer similar to those in a plane channel. Our findings demonstrate the robustness and limitations of the present method, highlighting the potential of defiltering for complex geometries in engineering applications.