JK-FLOW Japan-Korea Fluid Mechanics Online Workshop

Speaker: Dr. Mario Rüttgers (Researcher, Walter Benjamin Fellow, Inha University) [GS]

Abstract: With rising concerns regarding global warming and energy security, there is an increasing demand for renewable energy sources. Recently, meteorology-dependent renewable urban energy resources, i.e., urban wind turbines or solar energy devices, play a more and more important role in helping cities in shifting to an energy self-sufficient or energy positive status. This talk presents a tool that is developed for optimizing the utilization of urban wind turbines. The tool is realized with an optimization algorithm that combines and locates horizontal and vertical turbines in urban planning scenarios, while receiving feedback from urban flow predictions combined with the turbines’ power curves. The flow predictions are done by a graph convolutional neural network (GCNN) that is trained with data from computational fluid dynamics (CFD) simulations of randomly defined urban flows. The GCNN is fed with two types of inputs, i.e., information about the topology of the urban area, and wind conditions at the boundaries. The network is tested with cutouts of real cities and boundary conditions from publicly available meteorological data. The tool assists city planners in finding the perfect number and locations for urban wind turbines.

Speaker: Mr. Jiyeon Kim (Ph.D. student, Yonsei University) [GS]

Abstract: Recent advances in deep learning (DL) have highlighted the potential of generative models, which learn unknown data distributions to generate new samples from noise. By incorporating conditional inputs, these models enable various applications, including dynamics prediction, where past fields are used to predict future states. While generative adversarial network (GAN)-based models have dominated the field, challenges such as scalability and training instability persist. Recently, diffusion probabilistic models (DMs) have emerged, offering the robustness of likelihood-based approaches and achieving performance comparable to or surpassing GANs. However, their computational cost is significantly higher, often orders of magnitude greater than GANs with similar performance. This talk presents the application of DM to 2D turbulence prediction, along with a comprehensive performance evaluation against various DL models, including a conditional GAN. Our findings show that DMs outperform others at lead times shorter than the Eulerian integral time scale but experience significant performance degradation at longer lead times. Ongoing efforts to extend DMs to 3D turbulence prediction will also be discussed.

Speaker: Dr. Ryo Araki (Assistant Professor, Tokyo University of Science)

Abstract: The small-scale universality of developed turbulence is often described as the small scales "forgetting" the macroscopic flow characteristics during the scale-local energy cascade. However, turbulence is inherently causal; for example, temporal fluctuations of small-scale quantities (such as the energy dissipation rate) exhibit a time-delayed correlation with large-scale quantities (such as the energy input rate). To reconcile this apparent paradox, we analysed high-Reynolds-number homogeneous and isotropic turbulence using information flux, a measure of how knowledge of a variable's current state reduces uncertainty about the future state of another variable in a dynamical system. Our analysis revealed a scale-local forward information transfer within the inertial range, accompanying the energy cascade. Furthermore, we examined the roles of different cascade mechanisms - vortex stretching (VS) and strain self-amplification (SSA) - in both energy and information transfer. Our findings indicate that these two transfers are governed by different mechanisms: the dominant energy cascade mechanism is not necessarily the most causal one, and vice versa.

Speaker: Mr. Jungjae Woo (Ph.D. student, Korea University)

Abstract: The integration of microbubble technology into fluid systems has opened new avenues for efficient and eco-friendly cleaning applications. This talk will present recent advancements in utilizing microbubble jets to enhance oil removal and energy efficiency in various cleaning applications. Microbubbles, characterized by their unique hydrodynamic properties and prolonged stability, have been shown to improve the removal of oil contaminants through enhanced jet instability, surface interactions, and turbulence intensification. To better understand the underlying mechanisms, the hydrodynamic characteristics of microbubble jets were analyzed, focusing on their influence on jet instability, velocity fields, and breakup dynamics. By incorporating microbubbles into conventional water jets, cleaning efficiency can be increased while reducing the reliance on chemical detergents and excessive water usage. These findings highlight the role of microbubble-driven jet instabilities in modifying flow behavior and their potential in developing next-generation eco-friendly cleaning technologies with maximized performance and minimized energy consumption.

Speaker: Dr. Yelyn Ahn (Postdoctoral Research Associate, Seoul National University)

Abstract: TBD

Speaker: Dr. Ming Liu (Project Research Associate, The University of Tokyo) [GS]

Abstract: Turbulent simulations with wall models are commonly used approaches to produce high fidelity flow fields with acceptable computational cost. However, most existing wall models are built under certain assumptions, which can affect their adaptivity to practical turbulent flows. To this end, we develop a novel wall model based on a deep neural network, namely, discriminator, which can discriminate instantaneous under-resolved and well-resolved flow fields. The fully developed velocity fields from direct numerical simulations (DNSs) on fine and coarse grids are performed and then adopted as the datasets to train the discriminator. Then, the well-trained discriminator is implemented into DNSs on coarse grids as a wall model. This dynamically updating the instantaneous velocity fields so as to make them indistinguishable from well-resolved ones through body force. The turbulent flow under bulk Reynolds number of 4600-40000 are investigated. As the discriminator-based wall model is introduced, the predicted wall shear stress, mean and rms velocity profiles are significantly improved compared with DNSs on coarse grids without a wall model.

Speaker: Dr. Vedasri Godovarthi (Postdoctoral Research Associate, Johns Hopkins University) [GS]

Abstract: TBD