JK-FLOW Japan-Korea Fluid Mechanics Online Workshop

Speaker: Dr. Shaghayegh Zamani Ashtiani (Postdoc Research Associate, Tohoku University)

Author list: Shaghayegh Zamani Ashtiani, Soshi Kawai, and Kai Fukami

• Department of Aerospace Engineering, Tohoku University 

Abstract: Many unsteady aerodynamic flows appear complicated with a range of spatiotemporal length scales, yet the dynamics that govern them are often far simpler. Reduced-order models can extract this simplicity, but their design depends on the flow physics of interest. This talk centers on two-dimensional transonic buffet, a flow described by millions of degrees of freedom whose periodic nature leads to low-dimensional underlying dynamics. We compress the flow with proper orthogonal decomposition followed by a nonlinear autoencoder, whose structured latent space yields a three-dimensional embedding encoding buffet phase, frequency, and angle of attack. Inspired by the supercritical Hopf bifurcation, this embedding supports a parametrized reduced representation for quasi-steady pitching as well as flow-field reconstruction from sparse sensors. Towards the end of the talk, we also briefly introduce our recent efforts on time-dependent basis-based modal analysis that linearly evolves the bases directly in time rather than fixing them as in conventional approaches with an example of extreme vortex gust encounter. These complementary strategies show that encoding prior knowledge of flow physics in the model design is key to parametrization and, ultimately, surrogate modeling of unsteady aerodynamic flows.

Speaker: Mr. Hanbi Kang (Ph.D. student, Dong-A University)

Abstract: Computational fluid dynamics has become an essential tool for solving a wide range of engineering problems. However, the substantial cost of repetitive high-fidelity simulations limits its use in applications requiring rapid or long-term predictions. This challenge has motivated the development of reduced-order models (ROMs), which capture the dominant dynamics of high-dimensional systems in low-dimensional spaces. In particular, Koopman operator theory provides a promising framework for representing nonlinear dynamics through linear evolution in a space of observable functions. Building on this framework, Koopman autoencoders (KAEs) use neural networks to learn nonlinear latent representations while modeling their temporal evolution with a linear Koopman matrix. Despite their potential, conventional KAEs may suffer from accumulated errors and unstable long-term predictions when the spectral properties of the Koopman matrix are not properly constrained. In this work, we present KAEs that impose structural constraints on the latent dynamics. Using unsteady flow-field data, we compare the proposed approaches with a conventional KAE and examine how these constraints affect long-term prediction accuracy and stability. This presentation highlights the potential of structured latent dynamics for developing stable data-driven ROMs.

Speaker: Dr. Ye Won Kim (Postdoc Research Associate, Max Planck Institute)

Abstract: Shallow cumulus clouds over tropical oceans play a fundamental role in the Earth’s energy budget, yet the microphysical processes linking cloud droplets to precipitation remain poorly understood, particularly the interplay between turbulence, entrainment, and rapid droplet growth. Using high-resolution in-situ measurements from the EUREC⁴A campaign (tropical Atlantic, Jan–Feb 2020), we investigate how droplet size distributions relate to their centimeter-scale spatial organization within shallow cumulus clouds. By combining planar Particle Image Velocimetry (PIV) and holographic imaging, we simultaneously characterize local turbulence and three-dimensional droplet fields. Our analysis reveals that droplet clustering coincides with enhanced local turbulence, and that droplet size distributions differ systematically between clustered and void regions. These findings provide insights into how small-scale turbulence and spatial organization influence droplet growth pathways and contribute to precipitation formation in shallow cumulus clouds.

Speaker: Mr. Takuya Takase (Graduate student, Science Tokyo)

Abstract: Accurate prediction of urban wind fields is important for disaster mitigation and safe drone operations, but building-resolving simulations are too computationally expensive for real-time use. Existing super-resolution approaches can accelerate high-resolution flow prediction, yet still rely on building-resolving simulations. We propose a convolutional-neural-network-based three-dimensional super-resolution (3DSR) method that reconstructs building-scale temperature and velocity fields directly from mesoscale meteorological fields. The model is trained and evaluated using building-resolving large-eddy simulation (LES) data, and its transferability to a different urban area is examined using limited-data fine-tuning. The proposed method achieves temperature and velocity inference errors comparable to those reported in previous studies, while delivering a substantially greater speedup of approximately 680-fold relative to high-resolution LES. Fine-tuning improves accuracy in unseen urban areas, demonstrating the potential of 3DSR for fast and transferable urban micro-meteorological prediction.