Direct Numerical Simulation Of Flow Research Articles

Effects of microscopic pore-throat structure on gas–liquid relative permeability: Porous media construction and pore-scale simulation

The porous media structure of the oil/gas reservoir changes greatly after long-term development, which subsequently influences the macroscopic relative permeability. To clarify the effects of microscopic pore-throat structure on macroscopic relative permeability, we first proposed a method to generate two-dimensional porous media images with adjustable structure parameters. The method is based on Delaunay triangulation and is similar to the pore-network generation process, which can provide binary images for direct numerical simulation of flow through porous media. Then, we established the single component multiphase Shan–Chen lattice Boltzmann method coupling the real gas equation of state. Finally, we discussed the effect of pore radius, coordination number, pore-throat ratio, pore shape, and wettability on the gas–liquid relative permeability curve using the lattice Boltzmann simulation. This study provides an effective method to generate porous media and explain the mechanism of relative permeability change at the pore scale.

An analysis of drag reduction using spanwise forcing on rough walls

Spanwise opposed wall-jet forcing has been shown to reduce the skin-friction drag of wall-bounded turbulent flows by suppressing the near-wall turbulent motion (Yao et al., 2018). In the present work, the response of this drag reduction mechanism to the presence of surface roughness is studied. To this end, direct numerical simulations of flow in smooth and rough plane channels at a matched friction Reynolds number (Reτ=180) are carried out. The roughness is generated by distributing discrete roughness elements at two different values of frontal solidity, namely 0.021 and 0.045. The roughness elements height is k+=18 (＋indicates viscous scaling). By varying the forcing amplitude A+, it is shown that, when other forcing parameters are fixed, maximum drag reduction in a rough channel is achieved at a considerably larger A+ than in a smooth channel. Notably, the strength of the wall-jet at which the maximum drag reduction is achieved, is comparable for the smooth and both rough cases. Additionally, it is observed that the maximum drag reduction values attained in rough channels are smaller compared to those observed in smooth channels, for both lower and higher frontal solidity cases — 2.4% and 2%, respectively. The reduced drag reduction potential originates from two observations; firstly, drag on roughness elements, which is dominated by pressure drag, is not reduced by the method in question. Secondly, the drag on channel floor is reduced to a lesser extent when roughness elements are present. The latter can be attributed to the observation that, while in all cases the forcing suppresses streamwise random turbulent fluctuations and shear stress, this suppression is less pronounced when roughness elements are present.

Direct numerical simulation of flow and heat transfer of supercritical water with different heat fluxes

The flow and heat transfer characteristics of supercritical water (SCW) are important considerations in the thermal-hydraulic design of the Supercritical Water-Cooled Reactor (SCWR). In the present study, the direct numerical simulation (DNS), developed with OpenFOAM, was utilized to investigate the effect of heat flux on the flow and heat transfer of SCW in a vertical circular pipe. The investigation involves detailed analyses of the wall temperature, velocity distribution, and turbulent statistics at various heat fluxes. It was found that buoyancy first impairs heat transfer, and then recovers heat transfer. This is attributed to the variation of buoyancy production and turbulence production, which are resulted from the buoyancy direct effect and indirect effect, respectively. As the heat flux increases, the recovery occurs earlier with higher magnitude. The vortex structures based on Q-criterion clearly illustrate the same process as well. Several heat-transfer correlations were assessed based on the DNS results. It indicates that Dittus and Boelter correlation (1930) cannot accurately predict the heat-transfer characteristics of SCW, while Chen and Fang correlation (2014) exhibits a higher level of predictive accuracy for Nusselt number. In addition, several turbulence models were evaluated against the DNS data. The results show significant deviations in predicting the heat transfer of SCW. The large discrepancy may be due to the inappropriate model coefficients, the failed predication of turbulence kinetic energy, as well as the constant turbulent Prandtl number commonly applied.

Direct numerical simulation of flow in open rectangular ducts

We study turbulent flow in open channels with a free surface and rectangular cross-section, for various Reynolds numbers and duct aspect ratios. Direct numerical simulations are used to obtain accurate characterization of the secondary motions, which are found to be more intense than in closed ducts, and to scale with the bulk, rather than with the friction velocity. A notable feature of open-duct flows is the presence of a velocity dip, namely the peak velocity is achieved at some depth underneath the free surface. We find that the depth of the velocity peak increases with the Reynolds number, and correspondingly the flow becomes more symmetric with respect to the horizontal midplane. This is also confirmed from the change of the topology of the secondary motions, which exhibit a strong corner circulation at the free-surface/wall corners at low Reynolds number, which, however, weakens at higher $Re$ . The structure of the mean velocity field is such that the log law applies with good approximation in the direction normal to the nearest wall, which allows us to explain why predictive friction formulae based on the hydraulic diameter concept are successful. Additional analysis shows that the secondary motions account for a large fraction of the frictional drag (up to $15$ %).

Optimizing flow control with deep reinforcement learning: Plasma actuator placement around a square cylinder

This study introduces a deep reinforcement learning-based flow control approach to enhance the efficiency of multiple plasma actuators on a square cylinder. The research seeks to adjust the control inputs of these actuators to diminish both drag and lift forces on the cylinder, ensuring flow stability in the process. The proposed model uses a two-dimensional direct numerical simulation of flow past a square cylinder to represent the environment. The control approach involves adjusting the AC voltage across three specific configurations of the plasma actuators. Initially tested at a Reynolds number (ReD) of 100, this strategy was later applied at ReD of 180. We observed a 97% reduction in the mean drag coefficient at ReD = 100 and a 99% reduction at ReD = 180. Furthermore, the findings suggest that increasing the Reynolds number makes it harder to mitigate vortex shedding using plasma actuators on just the cylinder's rear surface. However, an optimized configuration of these actuators can fully suppress vortex shedding under the proposed control scheme.

A new flow-based design for double-lumen needles

Oocyte retrieval forms a crucial part of in vitro fertilisation treatment and its ultimate outcome. Standard double-lumen needles, which include a sequence of aspiration and flushing steps, are characterised by a similar success rate to single-lumen needles, despite their increased cost. A novel hydrodynamics-based needle called the OxIVF needle is proposed here, which is geared towards the generation of an internal flow field within the full follicular volume via laterally, rather than frontally, oriented flushing, leading to successful retrievals with no additional stress on the oocyte. A two-dimensional digital twin of the follicular environment is created and tested via multi-phase flow direct numerical simulation. Oocyte initial location within the follicle is varied, while quantities of interest such as velocity magnitude and vorticity are measured with a high level of precision. This provides insight into the overall fluid motion, as well as the trajectory and stresses experienced by the oocyte. A comparative benchmark set of tests indicated a higher success rate of the OxIVF needle of up to 100%, marking a significant improvement over the traditional double-lumen design whose success rate of no more than 75% was also highly dependent on the location of the needle tip inside the follicle. All forces measured during these tests showcase how the oocyte experiences stresses which are no larger than at the aspiration point, with the flow field providing a gentle steering effect towards the extraction region. Finally, the flow generation strategy maximises oocyte yield, unlocking new capabilities in both human and veterinary contexts.

Calibration of a Near-Wall Differential Reynolds Stress Model Using the Updated Direct Numerical Simulation Data and Its Assessment

In the article, a differential Reynolds stress model is recalibrated using turbulent channel flow direct numerical simulation data in the range of friction Reynolds numbers 550–5200. The calibration aims to produce a RANS sublayer model for use within the hybrid RANS/LES framework. The model is designed to capture the average field of a thin near-wall part of a boundary layer as accurately as possible. An a posteriori procedure is employed in which one-dimensional channel flow calculations are performed for all variations of the model coefficients at each stage of the optimization procedure. The coefficients are initialized with their original values and then optimized by minimizing the appropriately chosen norm. An improved representation of the mean velocity profile and peak Reynolds stress values is demonstrated. Both models—baseline and recalibrated—are implemented in an in-house CFD code, and several simulations, including a channel flow, a flat plate boundary layer and a boundary layer separation from a rounded step, are performed. The latter benchmark flow is also simulated in hybrid RANS/LES mode. The updated model is compared to the original one, demonstrating improvements over the baseline model in the cases it was designed for.

Direct numerical simulations of flows around an array of rough cylinders partially coated with porous media at Re = 3900

In recent years, interest has been growing in the stability control of cylindrical structures in a flow field using full/partial coatings of porous media. However, the partial application of porous coatings to arrays of cylinders with rough surfaces has rarely been studied. This paper numerically investigates the passive flow control of an array of rough cylinders partially coated with porous media on their leeward side in the subcritical flow regime (Re = 3900). This is achieved through two-dimensional direct numerical simulations and the Cartesian cut-cell method, which facilitate flow simulations around complex porous media and rough surfaces in an accurate and flexible manner. In contrast to macroscopic models, this study employs a quasi-microscopic model to simplify the porous structure for each cylinder, providing an accurate and intuitive means of describing the inflow details of porous media near the central cylinder. The porosities and pores per inch of the porous media are defined in terms of the alignment angles and radii. Similarly, the surface roughness of a cylinder is simulated by the attachment of small circular appendages. The accuracy and capacity of the presented numerical approach are demonstrated via a numerical case study with a single smooth cylinder. The influence of the porous media on the flows around a single rough cylinder and an array of rough cylinders is thoroughly investigated and discussed.

Validation of RANS-Based Turbulence Models Against High-Resolution Experiments and DNS for Buoyancy-Driven Flow with Stratified Fronts

To provide computational fluid dynamics (CFD)–grade experimental data for studying stratification, measurements on the High-Resolution Jet (HiRJet) facility at the University of Michigan have been conducted with density differences of and , respectively. Fluid with a density different from the fluid initially present in the HiRJet tank was injected, and the propagation of the time-dependent density stratification was captured on a two-dimensional plane with the aid of the wire-mesh sensor technique for Reynolds numbers near 5000 and Richardson numbers near 0.29. Direct numerical simulations (DNSs) of the two cases have also been conducted to expand the multifidelity database. The novel experimental and DNS data were then used to assess the predictive capabilities of the Standard (SKE) model and the Reynolds Stress Transport (RST) model. In particular, the propagation speed and thickness of the stratification fronts were assessed by comparing the CFD results against the experimental and DNS data. It was found that the general trends of the stratified density fronts were well predicted by the CFD simulations; however, slight overprediction of the thickness of the stratification layer was found with the SKE model while the RST model gave a larger overprediction of the mixing. Examination of the turbulent statistics showed that the turbulent viscosity was largely overpredicted by the RST model compared to the SKE model.

Design of a Direct numerical Simulation of flow and heat transfer in a T-junction

Several investigations have been undertaken to study the velocity and temperature fields associated with the thermal mixing between fluids, and resulting thermal striping in a T-junction. However, the available experimental databases are not sufficient to describe the involved physics in adequate detail, and, due to experimental limitations, accurate data on velocity and temperature fluctuations in regions close to the wall are not available. Computational Fluid Dynamics (CFD) can play an important role in predicting such complex flow features. However, predicting complex thermal fatigue phenomena is a challenge for the available momentum and heat flux turbulence models. Furthermore, such models need to be extensively validated.The aim of the present work is to design a reference numerical experiment for Direct Numerical Simulation (DNS) of a thermal fatigue scenario using Reynolds-Averaged Navier-Stokes (RANS) simulations. First, the feasibility of scaling down the Reynolds number from experimental cases to a computationally-feasible range is investigated. The junction corner shape is also modified to a slightly rounded corner, ensuring that the underlying fundamental physical phenomena of turbulence and thermal mixing flow features are preserved. Finally, the pipe lengths of the model were calibrated to ensure there would be no interference of the upstream developing region and the outlet boundary conditions on the thermal mixing at the junction. A sample under-resolved DNS case, with unity and low-Prandtl number passive temperature scalars, with iso-temperature, iso-flux and mixed (Robin) wall boundary conditions, are presented. This proof-of-concept simulation contributes to the finalization of the set-up for fully-resolved DNS with respect to the computational grid size selection and transient characteristics.

Direct numerical simulation of flow past a reactive/inert mixed particle pair

During the pyrolysis and gasification processes inside fluidized bed reactors, a single reactive particle is generally surrounded by multiple inert heat carrier particles. In this work, we focus on the simplest mixed particle pair (one reactive particle: P1, and one inert particle: P2) to investigate the joint effects of the Stefan Reynolds number (Resf), the particle Reynolds number (Re), separation distance normalized by particle diameter (L), and relative orientation (α) on the drag force exerting on each particle by particle-resolved direct numerical simulations. The results show that when particles are placed in a tandem arrangement, the blocking effect on P2 is weakened/strengthened with negative/positive Stefan flow since the effective diameter of P1 is smaller/larger than its real diameter. When L = 1.25, Re = 3, and particles are placed in tandem, the drag force of P1 decreases as Resf increases, but the reduction of the drag force decreases as Re increases. Under such conditions, the positive Stefan flow induces the repulsion between the two particles; thus, the drag force of P2 increases as Resf increases. When Resf = 3 and P2 is located downstream, the repulsive force felt by P1 can partially offset the drag force at L = 1.25, Re = 3. Moreover, the negative/positive Stefan flow enhances/weakens the drag force of P2 when it is located upstream. The Stefan flow has a negligible effect on the drag force of P2 when L = 3.

Direct numerical simulation of flow and heat transport in a closely-spaced bare rod bundle

Direct Numerical Simulation (DNS) of fully-developed flow and heat transfer is performed for a bare rod bundle with a low pitch-to-diameter ratio using the highly-scalabale spectral element code Nek5000. The simulation is performed at a Reynolds number, Reh, based on the bulk velocity and hydraulic diameter, of 9800, with iso-temperature and iso-flux thermal boundary conditions and Prandtl numbers 0.025, 1.0 and 2.0. The mean velocity and temperature statistics are first validated against the experimental measurements of Hooper et al. (1983) and complementary DNS results of Lai et al. (2019) which are reported in the literature at a higher Reh of 22,600. The mean flow and temperature are demonstrated a similar behaviour at the present low Renolds number and the higher Reynolds number of Lai et al. (2019), indicating the Reynolds number scaling of the mean flow. The anisotropy of the flow is illustrated using distribution of normal and Reynolds stresses, and the Lumley invariant map. It is shown that the turbulence in the subchannel region is characteristic of wall-bounded turbulent flows such as channel flows. In the narrow gap, however, turbulence is suppressed, only enhanced in the streamwise direction due to flow pulsations. Further, temperature statistics with both thermal boundary conditions are presented, demonstrating the effect of the Prandtl number on the mean and fluctuating quantities. The temperature statistics with iso-temperature wall boundary conditions, not reported in literature before, are shown to behave differently in the interstitial subchannel region and the narrow gap between two rods. The distribution of turbulent heat flux is also presented. The streamwise component of turbulent heat flux is again shown to be enhanced, in comparison to the same in the subchannel. The pulsating motion of flow across the narrow gap is discussed further. A frequency analysis of the velocity in the narrow gap elucidates the frequency of flow pulsations. Unlike mean flow statistics, the flow pulsations are shown to exhibit a low Reynolds number effect. The reference DNS data presented in the present work, complementary to that of Lai et al. (2019), serves as a benchmark against which lower order turbulence models may be further developed or improved. In particular, the newly presented anisotropy data and the turbulence heat flux data are essential for further development, improvement and validation of turbulent heat transfer models at different Prandtl numbers.

Status, perspectives, and added value of high fidelity simulations for safety and design

With the increase in computational power, High Fidelity (HiFi) simulations become increasingly important to nuclear thermal hydraulics for fundamental understanding of turbulent flow and heat transfer phenomena and for further development and validation of engineering CFD models. We consider Direct Numerical Simulations (DNS) and Large Eddy Simulation (LES) as HiFi analyses. In this paper, we present an overview of single-phase and two-phase flow DNS simulations relevant for nuclear reactor safety and design purposes. In order to limit the scope, this overview has a focus on DNS simulations of fully turbulent non-reacting incompressible flows. LES approaches are briefly mentioned in certain areas, where LES-to-DNS upgrade is expected to happen in the near future.We first explain the concept of DNS and related concepts like Under-resolved DNS (UDNS) and quasi-DNS (q-DNS). Subsequently, we present the numerical methods which are generally used for HiFi simulations. A wider span of numerical methods can be observed in single-phase flow simulations than in two-phase flow simulations, where the later require additional algorithms for tracking of the interfaces between the two phases. Consequently, two-phase DNS simulations are not immune to modelling errors, and are much more complex than their single-phase counterparts. Next, an overview is given of single-phase flow DNS simulations. We start with basic cases like turbulent channel and pipe flows, which are mainly useful for development and validation of less detailed engineering CFD models. We continue with more recent and more complex cases like, e.g., flows in a rod bundle and a pebble bed, which are directly applicable in nuclear engineering. Next, an overview of two-phase flow HiFi simulations is presented in a similar fashion. The importance of the availability of the presented HiFi simulations is explained by addressing their role for further development and validation of engineering CFD models which are used for the eventual industrial applications. The importance of HiFi simulation data, complementary to experimental data, for uncertainty quantification and machine learning is explained. Finally, future challenges for HiFi simulations, especially for two-phase flow, are elucidated.

Data-Driven Unsteady Aeroelastic Modeling for Control

Aeroelastic structures, from insect wings to wind turbine blades, experience transient unsteady aerodynamic loads that are coupled to their motion. Effective real-time control of flexible structures relies on accurate and efficient predictions of both the unsteady aeroelastic forces and airfoil deformation. For rigid wings, classical unsteady aerodynamic models have recently been reformulated in state space for control and extended to include viscous effects. Here, we further extend this modeling framework to include the deformation of a flexible wing in addition to the quasi-steady, added mass, and unsteady viscous forces. We develop low-order linear models based on data from direct numerical simulations of flow past a flexible wing at a low Reynolds number. We demonstrate the effectiveness of these models to track aggressive maneuvers with model predictive control while constraining maximum wing deformation. This system identification approach provides an interpretable, accurate, and low-dimensional representation of an aeroelastic system that can aid in system and controller design for applications where transients play an important role.

Direct numerical simulation of transitional flow past an airfoil with partially covered wavy roughness elements

We perform direct numerical simulation of flow past the NACA (National Advisory Committee for Aeronautics) 0012 airfoil with partially covered wavy roughness elements near the leading edge. The Reynolds number Rec based on the freestream velocity (U0) and airfoil chord length (C), and the angle of attack (AoA) is fixed, i.e., Rec=5×104, AoA=10°. The leading edge roughness elements are characterized by streamwise sinusoidal-wave geometry that is uniform in the spanwise direction with fixed height, whereas varying wave numbers (k) from 0 to 12. Based on the validation of the smooth baseline case (k = 0), the roughness effects on the aerodynamic performance are evaluated in terms of the lift and drag coefficients. The drag coefficient decreases monotonically with k, while the variation of the lift coefficient with k is similar to the “drag crisis” phenomenon observed in cylinder flows. The sharp variations of lift and drag coefficients from k = 6 to 8 indicate that k = 8 is a critical case, beyond which massive separation occurs and almost covers the airfoil's suction side and dominates the airfoil aerodynamic performance. To further reveal the underlying mechanism, the flow structures, pressure, skin friction coefficients, shear layer transition onset, and boundary layer separation are quantified to investigate the roughness effects. The roughness elements strongly modify the separation behavior, whereas they have little effect on the transition onset. The unsteady interactions and convections of separation bubbles downward the trailing edge also change the wake evolution. Based on the scaling of the asymmetric wake profiles, we find that the wake defect is gently decreasing with k, but the increase in the wake width is much stronger. This scaling analysis gains a better insight into the roughness effects on the momentum deficit flow rate, which confirms the drag mechanism with different roughness wave numbers.

Microscale drag model considering the effect of interface between dense and dilute phases for gas–solid suspensions at moderate Reynolds numbers

Particle-resolved direct numerical simulations of flows past various fixed structures of spheres are performed to study the drag force at local Reynolds numbers up to 250. At the gas–solid interface, the drag force is found to be 5%∼800% of that estimated by BVK law (Beetstra et al. AIChE Journal, 2007, 53(2):489-501) at different particle Reynolds numbers and solid volume fraction gradients. Thus a microscale drag model at the cluster interface is developed as a function of the particle Reynolds number and the solid concentrations of dense and dilute phases. Assessments of the proposed drag model are performed with direct numerical simulations of flows past plug-like, spherical, ellipsoidal clusters as well as bubble-containing structures. The results show that the new model well captures the drag variations at microscales of several sphere diameters and correlates well with the solid concentrations and velocities of dense and dilute phases.

Vertical and torsional vibrations before the collapse of the Tacoma Narrows Bridge in 1940

We perform a three-dimensional direct numerical simulation of flow over the Tacoma Narrows Bridge to understand the vertical and torsional vibrations that occurred before its collapse in 1940. Real-scale structural parameters of the bridge are used for the simulation. The Reynolds number based on the free-stream velocity and height of the deck fence is lower ( ${Re}=10\ 000$ ) than the actual one on the day of its collapse ( ${Re}=3.06 \times 10^{6}$ ), but the magnitude of a fluid property is modified to provide the real-scale aerodynamic force and moment on the deck. The vertical and torsional vibrations are simulated through two-way coupling of the fluid flow and structural motion. The vertical vibration occurs from the frequency lock-in with the vortex shedding, and its wavelength and frequency agree well with the recorded data in 1940. After saturation of the vertical vibration, a torsional vibration resulting from the aeroelastic fluttering grows exponentially in time, with its wavelength and frequency in excellent agreement with the recorded data of the incident. The critical flutter wind speed for the growth of torsional vibration is obtained as $3.56 < U_c / (f_{nat} B) \le 4$ , where $U_{c}$ is the critical flutter wind speed, $f_{nat}$ is the natural frequency of the torsional vibration and $B$ is the deck width. Finally, apart from the actual vibration process in 1940, we perform more numerical simulations to investigate the roles of the free-stream velocity and vertical vibration in the growth of the torsional vibration.

Pore-scale direct numerical simulation of fluid dynamics, conduction and convection heat transfer in open-cell Voronoi porous foams

The pore-scale analysis includes the pore network simulation and direct numerical simulation, which the latter is more accurate and provides more details about the problem. This study presents a pore-scale direct numerical simulation of flow and convection-conduction heat transfer in open-cell metal foams with a complex structure. First, using the Laguerre–Voronoi tessellations (LVT) algorithm, three foam samples with various porosities (76.6%, 84.4%, and 93.8%) and a constant pore density of 20 pores per inch (PPI) are generated. Next, the open-source library of OpenFOAM is employed to perform heat transfer and flow analysis in Darcy and non-Darcy regimes. Results were verified through comparison with available experimental and numerical data. The effects of foams' porosity, solid material, and Reynolds number on flow characteristics and heat transfer are investigated. Results are presented in terms of the permeability, Forchheimer coefficient, the volumetric and average heat transfer coefficient, and temperature and velocity distribution. By decreasing porosity, the flow complexity increases with subsequent increments in heat transfer and pressure drop. Also, the conductive heat transfer through the foam plays a vital role in the total heat transfer rate. These pore-scale outcomes could be used in macro-scale thermo-fluid models to perform more accurate simulations with less computational costs.

Riblet Drag Reduction Modeling and Simulation

One of the most interesting passive drag reduction techniques is based on the use of riblets or streamwise grooved surfaces. Detailed flow features inside the grooves can be numerically detected only by Direct Numerical Simulations (DNS), still unfeasible for high Reynolds numbers and complex flows. Many papers report the DNS of flows on microgrooved surfaces providing fundamental details on the drag reduction devices, but all are limited to plate or channel flows far from engineering Reynolds numbers. The numerical simulation of riblets and other drag reduction devices at very high Reynolds numbers is difficult to perform due to the riblet dimensions (microns in aeronautical applications). To overcome these difficulties, some models for riblet simulation have been developed in recent years, due to the data provided by DNS, experiments, and theoretical analyses. In all these models, the drag reduction is modeled rather than effectively captured; however, the analysis of some nonlocal effects on practical aeronautical configurations with riblets, requires their adoption. In this paper, the capabilities of these models in predicting riblets’ performance and some interesting features of the riblets’ effect on form drag and shock waves are shown. Two models are discussed and compared showing their respective advantages and limitations and providing possible enhancements. A comparison between the two models in terms of accuracy and convergence is discussed, and two new formulae are proposed to improve one of these models. Finally, a review of the results obtained by the two models is provided showing their capabilities in the analysis of the riblet effect on complex configurations.

Deep reinforcement learning for active control of flow over a circular cylinder with rotational oscillations

An active control strategy for flow over a circular cylinder with rotational oscillations is automatically learned based on a deep reinforcement learning algorithm. Specifically, the proximal policy optimization is used to minimize the averaged drag coefficient of cylinder by properly selecting the rotational angular velocity. The data for training is sampled by direct numerical simulation of flow past a rotational circular cylinder, which is recognized by the immersed boundary method and the active control is applied to the flow in real time. The flow configurations are observed by the detecting probes located inside them and modified by the agent of deep reinforcement learning through changing the angular velocity with time. After several episodes of training, the reinforcement learning agent is shown to discover a robust and convergent control strategy, by which the cylinder is able to choose an appropriate rotational angular velocity to modify the wake based on its observation of the instantaneous flow field, resulting in a substantial drag reduction. The network architecture is tested at different Reynolds numbers, i.e. 100, 200 and 300, and drag reduction rates of 10%, 19% and 21% are obtained, respectively. Effects of initial flow field and angular velocity, position and number of detecting probes and hyper-parameters of the deep reinforcement learning architecture are also examined.