Flow-induced noise during the charging of hydrogen fuel for a fuel cell electric vehicle is numerically analyzed. The noise over the entire storage system is predicted by solving the unsteady version of the Reynolds-averaged Navier–Stokes equations and the Ffowcs Williams–Hawkings acoustic analogy. The accuracy of the simulation results is validated by comparing pressure, temperature, and sound pressure levels with experimental data. The contribution of each component to total noise is investigated by separately setting the integration surfaces of the acoustic analysis, and it is found that the receptacle produces the largest noise level. The large acoustic source terms, calculated by the total derivative of the pressure, occur where the pipe bends and at narrow passages of check valves. From the comparative analysis of different manifold branching configurations, a layout in which all the pipe junctions are concentrated in a single manifold leads to increased noise levels.

A deep-reinforcement-learning (DRL) based control method to take the advantage of complex wake interactions in a wind farm is developed. Although the wind over a wind farm is changing, steady wind has been assumed in the most conventional methods for wind farm control. Under unsteady wind, the generated power of a wind farm becomes stochastic due to intermittent and fluctuating wind. To tackle the difficulty, a DRL-based method with which the pitch and yaw angles of wind turbines in a wind farm are strategically controlled is developed. Time-histories of the past wind and the predicted future wind are both utilized to identify the relation between the generated power and control. The present neural network is trained and validated using an experimental wind farm. A multi-fan wind tunnel is developed to generate unsteady wind for experiments with miniature wind farms, where the improvement in the generated power by the present DRL-based control method is demonstrated.

A comprehensive review of recent advancements in applying deep reinforcement learning (DRL) to fluid dynamics problems is presented. Applications in flow control and shape optimization, the primary fields where DRL is currently utilized, are thoroughly examined. Moreover, the review introduces emerging research trends in automation within computational fluid dynamics, a promising field for enhancing the efficiency and reliability of numerical analysis. Emphasis is placed on strategies developed to overcome challenges in applying DRL to complex, real-world engineering problems, such as data efficiency, turbulence, and partial observability. Specifically, the implementations of transfer learning, multi-agent reinforcement learning, and the partially observable Markov decision process are discussed, illustrating how these techniques can provide solutions to such issues. Finally, future research directions that could further advance the integration of DRL in fluid dynamics research are highlighted.

A new turbulence model has been developed for a Reynolds-averaged Navier–Stokes (RANS) simulations of buoyancy-driven flows. This study proposes a modification to the buoyancy-related term in the conventional k–ε RANS model's ε equation. Typical two-equation RANS models provide accurate predictions in homogeneous shear flow, decaying turbulence, and log-law regions, but have uncertain effectiveness for buoyancy-driven flows, particularly concerning the buoyancy-related term in the ε equation. They have produced significant errors in natural convection scenarios where the buoyancy-related term dominantly affects the modeling results, such as in the Rayleigh–Bénard (RB) convection. Conventional models are known to inaccurately predict RB convection when treated as a steady-state problem with zero mean velocity, considering only the gravity-directed coordinate as the independent variable. The analysis reveals that the conventional RANS model, along with modeling terms for buoyancy effects, provides not only inaccurate but also divergent turbulent heat fluxes in RB convection at high Rayleigh numbers. The proposed model establishes mathematical conditions that enable steady-state RANS simulations to converge to consistent scaling relations for the Nusselt number across a wide range of Rayleigh and Prandtl numbers in RB convection. This approach algebraically modifies a single term in the ε equation, so that the term vanishes in the absence of buoyancy, so the modification integrates seamlessly with the conventional k–ε RANS model.

A multi-objective optimization method utilizing genetic algorithms is developed to optimize wind farm layout design with the dual objectives of enhancing the production of wind power and reducing hour-level intermittency in the generated power. This method introduces a novel metric for annual wind power intermittency based on the transition probability of wind conditions (i.e., speed and direction). The present multi-objective optimization method ensures that Pareto optimal layouts are evenly distributed according to multiple cost function values. To mitigate wind power intermittency, the spatial distribution of wake fields is strategically manipulated to counteract hourly fluctuations in wind conditions. In wind conditions favoring high power generation, it is found that increasing the overlap between wakes and turbines helps to minimize disparities in generated power compared to other wind conditions.

A method using deep reinforcement learning (DRL) to non-iteratively generate an optimal mesh for an arbitrary blade passage is developed. Despite automation in mesh generation using either an empirical approach or an optimization algorithm, repeated tuning of meshing parameters is still required for a new geometry. The method developed herein employs a DRL-based multi-condition optimization technique to define optimal meshing parameters as a function of the blade geometry, attaining automation, minimization of human intervention, and computational efficiency. The meshing parameters are optimized by training an elliptic mesh generator which generates a structured mesh for a blade passage with an arbitrary blade geometry. During each episode of the DRL process, the mesh generator is trained to produce an optimal mesh for a randomly selected blade passage by updating the meshing parameters until the mesh quality, as measured by the ratio of determinants of the Jacobian matrices and the skewness, reaches the highest level. Once the training is completed, the mesh generator creates an optimal mesh for a new arbitrary blade passage in a single try without an repetitive process for the parameter tuning for mesh generation from the scratch. The effectiveness and robustness of the proposed method are demonstrated through the generation of meshes for various blade passages.

Numerical simulations of turbulent flow require excessive computational resources due to the multi scale characteristics of turbulence. Thus, a technique to reconstruct a high-resolution flow field from a coarse flow data can be helpful. For this purpose, various artificial neural-network-based super-resolution methods have been developed in recent years. Although previous studies reported that the super-resolution methods show remarkable performance for turbulent channel flow and homogeneous isotropic turbulence, its application for spatially developing flow with laminar, transitional, and turbulent characteristics has not been reported. In the present study, a super-resolution reconstruction method applicable for spatially developing laminar-transition-turbulent flow is developed by training the network for boundary layer flow with bypass transition. In addition, the generalization of the network for boundary layer flow with natural transition and for fully turbulent boundary layer flow is attempted. A super-resolution method based on a generative adversarial network (GAN) is employed for the study as it shows the best performance among tested network models. It is found that the developed method successfully reconstructs flow structures in transitional and early turbulent regions. In addition, statistics such as the mean velocity and the power spectral density of velocity from recovered fields show good agreement to those of DNS. Notably, the GAN model which is only trained for the bypass transition is also found to be applicable to boundary layer flow with K-type natural transition and fully developed turbulent boundary layers.

In the initial design stage of axial gas turbines, calculation algorithms namely through-flow analysis methods are used. In the conventional through-flow analysis methods, separate algorithms have been selectively employed for the subsonic and supersonic flow regimes, which complicates the solution procedure and degrades the computational efficiency. In the present study, a new unified through-flow analysis algorithm which can treat both the subsonic and supersonic flow regimes is proposed. Unlike in the conventional algorithms, where two Mach number solutions are calculated at every calculation points of through flow, the pressure field is updated for each iteration such that a unique Mach number is determined at every calculation points of through flow in the present algorithm. The developed algorithm allows a through-flow analysis for both the subsonic and supersonic flow regimes without user intervention. Numerical simulations are conducted to validate the proposed method, and the results confirmed the prediction accuracy.

Heat transfer in a continuous casting mold is produced through interfacial heat fluxes from the molten steel to the mold hot-face and convection between the mold and water. Although conventional methods based on empirical correlations are advantageous in terms of computational cost, they are not capable of capturing detailed cooling effects. Although, computational fluid dynamics methods coupled with a heat transfer analysis can provide detailed and accurate distributions of the heat transfer phenomenon in the continuous casting mold, the geometric complexity and associated computational difficulties hinder the practical use. To overcome the difficulties, in the present study, the characteristics of the forced convection coefficient in a typical casting mold is firstly identified. Based on the understanding of the heat transfer characteristics, a partial model approach where the conjugate fluid and heat transfer analysis is conducted only for a selected portion of the mold, and which is capable of providing the forced convection coefficient accurately and efficiently is proposed. The proposed method is found to be capable of predicting the forced convection coefficient and the temperature with an error of less than 2 pct from those of a full model conjugate fluid-heat transfer analysis.

A computational study of the blood flow in a stenosed microvessel is presented using coupled discrete ghost-cell and continuous-forcing immersed boundary methods. This study focuses on studying platelet behaviors near the stenosis with deformable red blood cells (RBCs). The influence of varying hematocrit, area blockage, stenosis shape, and driving force on flow characteristics, RBCs, and particle behaviors is considered. Distinct flow characteristics are observed in stenosed microvessels in the presence of RBCs. The motion of RBCs is the major cause of time-dependent oscillations in flow rates, while the contribution of particles to the fluctuations is negligible. However, this effect decreases when the stenosis is elongated in the axial direction. Interestingly, as the hematocrit level increases, downstream particles move closer to the vessel wall due to the enhanced shear-induced lift force resulting from the interaction among RBCs and particles. Furthermore, it is observed that geometrical changes in the stenosis have a more significant impact on the axial profile of particle concentration compared to changes in hematocrit or driving force. An asymmetric stenosis leads to asymmetric profiles in the flow velocity and the distribution of cells and particles due to the geometric focusing effect of the stenosis. There is no significant change in flow rates until a blockage of 0%-50%, but a sudden increase in the root mean square of flow rates occurs at an 80% blockage. This study contributes to our understanding of the rheological behaviors of RBCs and rigid particles in a stenosed microvessel under various hemodynamic conditions.