JK-FLOW Japan-Korea Fluid Mechanics Online Workshop

Speaker: Dr. Pierluigi Morra (Postdoc Research Associate, Johns Hopkins University) [GS]

Abstract: The performance of hypersonic vehicles is sensitive to environmental disturbances, especially in transitional flow. Accurate and efficient prediction of the flow state from limited sensors is critical in both fundamental studies and applications. Recent work has shown that assimilating scarce data into direct numerical simulations (Buchta et al., JFM, 947, R2, 2022) can reconstruct full flow fields, but at high computational cost. The computational burden of high fidelity simulations hinders broad adoption, particularly for large experimental campaigns or practical use. Here, we introduce a deep-learning approach that accelerates assimilation by two orders of magnitude in terms of experiments processed per unit time. We minimize simulations by optimally sampling the solution space, use a deep operator network (DeepONet) as a proxy for the compressible Navier–Stokes equations, and apply a gradient-free search to efficiently identify optimal solutions. The method is demonstrated on the assimilation of wind-tunnel measurements in Mach 6 boundary-layer flow over a 7-degree half-angle cone.

Speaker: Dr. Yutaro Motoori (Assistant Professor, The University of Osaka) [GS]

Abstract: It is well known that vortices of various sizes coexist in turbulence. However, when we visualize vortices using vorticity or the second invariant of the velocity gradient tensor, only the smallest-scale vortices are prominent. To identify vortices at arbitrary scales, it is therefore necessary to decompose turbulence into different scales. As shown in the visualization, the scale decomposition reveals that various-size vortices form hierarchical structures. In the present study, we conduct direct numerical simulations of wall turbulence, such as turbulent boundary layers and channel flows, to examine the hierarchy of coherent vortices. Based on the hierarchy of vortices, we discuss the sustaining mechanism of turbulent boundary layers and channel flows, and clarify both the universality and dissimilarity between these two turbulent flows.

Speaker: Mr. Jihoon Kim (Ph.D. student, Korea University) [GS]

Author list: Jihoon Kim[1], Jeonglae Kim[2], Jaiyoung Ryu[1]

1. Department of Mechanical Engineering, Korea University, Seoul, Republic of Korea

2. School for Engineering of Matter, Transport and Energy, Arizona State University, Tempe, USA

Abstract: Shock wave/turbulent boundary layer interactions (SWTBLI) in supersonic regimes are critical to the aerodynamic characteristics of high-speed aircraft. Detailed understanding of the interactions facilitates the development of future supersonic and hypersonic vehicles. Direct numerical simulations (DNS) are performed to solve the compressible Navier-Stokes equations for describing SWTBLI over a 24° compression ramp at a freestream Mach number of 2.9. Fully developed turbulent flows are imposed at the inflow using a recycling-rescaling method with a recycling distance of 10δ_in. The reference station is selected based on the distance from the compression corner, which is 7δ_in. Taylor microscales and Kolmogorov lengths, and Reynolds numbers are evaluated at the reference station for three different boundary-layer thicknesses. Simulation results are validated for mean velocity, density-scaled root-mean-square velocity fluctuations, and two-point correlations. From the turbulent kinetic energy budget, the mechanism of turbulence amplification influenced by the boundary-layer thickness is discussed.

Speaker: Mr. Soju Maejima (Ph.D. student, Tohoku University) [GS]

Abstract: The use of very coarse computational grids for large-eddy simulations (LES) causes the resolved turbulence to significantly deviate from the physically accurate turbulence. This deviation inhibits the training for a machine-learning-based sub-grid scale (SGS) model, where supervised training with the filtered direct numerical simulation (fDNS) solution as the proxy for the LES solution is often employed. This study proposes the unsupervised-supervised machine-learning pipeline as an SGS model for very-coarse LES (vLES). The key part of the pipeline is the unsupervised CycleGAN, which enables the super-resolution of the nonphysical vLES flowfields. The predicted high-wavenumber components are then extracted as the SGS stresses. The a posteriori test using the turbulent channel flow shows that the proposed method results in the accurate prediction of the near-wall Reynolds shear stress and the resulting mean velocity profile. The budget analyses of the Reynolds stresses reveal that the proposed SGS model predicts significant SGS backscatter in the spanwise normal stress component in the near-wall region, and that it is crucial for the accurate prediction of the mean velocity.

Speaker: Mr.  Shilaj Baral (Ph.D. student, POSTECH) [GS]

Author list: Shilaj Baral [1], Youngkyu Lee [2], Sangam Khanal [1], and Joongoo Jeon [3]

1. Graduate School of Integrated Energy-AI, Jeonbuk National University, Republic of Korea 

2. Division of Applied Mathematics, Brown University, United States

3. Division of Advanced Nuclear Engineering, Pohang University of Science and Technology, South Korea

Abstract: The practical utility of hybrid methods for accelerating computational fluid dynamics (CFD) simulations hinges on their ability to generalize across diverse conditions and scale to complex, three-dimensional problems. This study investigates these critical properties using XRePIT, a novel, fully automated framework designed for this purpose. The framework couples OpenFOAM with machine learning surrogates, using a residual-guided loop to ensure long-term stability and physical accuracy. To assess generalization, we tested the hybrid method against multiple boundary conditions and demonstrated its architectural extensibility by seamlessly swapping between a finite-volume method network (FVMN) and a Fourier neural operator variant (FVFNO). To prove scalability, we extended the application from 2D benchmarks to a full 3D simulation of buoyancy-driven flow. Our results validate the approach, achieving stable accelerations of up to 3.68× in 2D and a significant 4.98× in 3D. Across all configurations, the method maintained long-term stability for over 10,000 timesteps with less than 1% error. This work provides critical evidence that the residual-guided hybrid strategy is not only a viable concept but a scalable and generalizable solution, marking a practical step towards applying ML-accelerated CFD to real-world, 3D engineering challenges.

Speaker: Dr. Guo Yuting (Assistant Professor, Kyoto University) [GS]

Abstract: TBA