Georgios Rigas

Controlling turbulent dynamics remains a major challenge because their chaotic, multiscale nature strongly affects many fluid systems. Here, we present REACT, an autonomous reinforcement learning framework for real-time, state-dependent control of turbulent wake dynamics in a wind tunnel. Deployed on an Ahmed-body model equipped only with onboard sensors and servo-actuated surfaces, REACT learns from sparse wind-tunnel measurements, bypassing empirical turbulence models. The agent autonomously converges to a policy that reduces aerodynamic drag while delivering net energy savings. Without prior knowledge of flow physics, it discovers that dynamically suppressing spatiotemporally coherent wake structures maximizes energy efficiency, achieving two to four times greater performance than model-based baseline controllers. We contrast REACT’s dynamics-aware, state-dependent policy with quasi-steady, mean-flow-oriented policies learned by standard reinforcement learning baselines, which yield lower drag reduction and do not suppress coherent instabilities in this turbulent-wake regime. Finally, by training in a nondimensional state-reward space with amplitudes approximately invariant to Reynolds number, and conditioning on Reynolds number for temporal adaptation, REACT learns an offline policy that remains effective across Re = 86,400-518,400 without retraining. These results demonstrate autonomous closed-loop reinforcement learning control in a high-Reynolds-number wind-tunnel environment and suggest a path toward data-driven, state-dependent control of turbulent flows.

We investigate the broadband turbulent dynamics of attached and separated flows over a Gaussian bump, focusing on the origin of low-frequency coherent structures. The analysis combines time-resolved experimental measurements with physics-based linear models, using mean fields previously assimilated from the same dataset as base flows. Spectral proper orthogonal decomposition reveals a coherent dynamics in low- and medium-frequency regimes for both flows, with the low-frequency dynamics being substantially stronger in the separated case. In the separated flow, this dynamics is linked to a three-dimensional zero-frequency modal instability that generates large-scale streamwise-elongated structures downstream of the bump. A standing-wave model based on resolvent modes, incorporating finite-span effects, reproduces the experimentally observed spanwise structure of the dynamics and highlights the limitations of simulations with small spanwise extent and periodic boundary conditions. In the attached flow, similar low-frequency structures are identified. These are weaker, do not form a prominent standing-wave pattern and cannot be definitively classified as either modal or non-modal. The three-dimensional zero-frequency instability and finite-span standing-wave dynamics are identified as the main drivers of low-frequency coherent structures in the separated flow. They offer an explanation for persistent discrepancies between simulations and experiments on the Gaussian bump, and provide guidance on spanwise domain size and boundary conditions for future simulations.

The extreme heat fluxes characteristic of hypersonic flows significantly limit the flight envelope of hypersonic vehicles. The role of hydrodynamic instability and the onset of laminar-to-turbulent boundary layer transition is of notable importance. The effect of streaks on the suppression of planar (second Mack mode) instabilities has been previously investigated, but a potentially passive and non-intrusive control method has not been established yet. Recent work shows that streaks can be generated through a spanwise variation in surface temperature. This method exploits the aerothermodynamic characteristics of the flow, and therefore promises to be robust. This work uses direct numerical simulations to determine and quantify the effectiveness of this novel control method in the suppression of second Mack mode instability for a hypersonic boundary layer over a flat plate. The computational analyses cover a range of Mach numbers, 4.8–6, and wall temperature ratios representative of both wind tunnel testing and flight scenarios. Among the range of configurations investigated, the energy of the second Mack mode is reduced by up to approximately 60 % by the steady streaks. The streak wavelength parameter plays a significant role in the stabilisation benefits. For a Mach 6 configuration, for the most linearly amplified second Mack mode disturbance frequency, nearly optimum performance is achieved for a spanwise wavelength of approximately 8–10 times the local boundary layer thickness. These findings open new avenues for controlling hypersonic boundary layers and offer valuable guidance for future experimental campaigns aimed at validating this novel control strategy.

Traditional wind farm control operates each turbine independently to maximize individual power output. However, coordinated wake steering across the entire farm can substantially increase the combined wind farm energy production. Although dynamic closed-loop control has proven effective in flow control applications, wind farm optimization has relied primarily on static, low-fidelity simulators that do not resolve critical dynamic turbulent fluctuations in the flow. In this work, we present a reinforcement learning controller trained using high-fidelity turbulence resolving simulations, enabling real-time response to atmospheric turbulence through collaborative, dynamic control strategies. In a three wind turbine test case, our reinforcement learning controller achieves a 4.30% (95% CI = [4.10%, 4.49%]) increase in wind farm power output compared to baseline operation, nearly doubling the 2.19% (95% CI = [1.98%, 2.39%]) gain from static optimal yaw control and a substantial increase over the gain from global wind direction based dynamic control obtained through Bayesian optimization of 2.67% (95% CI = [2.47%, 2.87%]). These results establish that reinforcement learning is able to utilize the increased information available from turbulence resolved simulations to learn improved, dynamic flow-responsive control for wind farm power maximization, with direct implications for accelerating renewable energy deployment to net-zero targets.

*[NOTE: the slides are not visible for the first five minutes of the talk! Audio was captured correctly for the whole duration of the talk.]* Reinforcement learning (RL) is increasingly used to discover control strategies for complex physical systems, yet turbulent flows remain a challenge: the dynamics are high-dimensional and nonlinear, sensing is sparse and noisy, high-fidelity simulations are computationally intractable, and experiments are expensive. In this talk, I will present recent work aimed at making RL practical for real-world flow control by combining sample-efficient learning with physics-aware representations of partially observable dynamics inherent to realistic sensor configurations.

Laminar-to-turbulent boundary-layer transition leads to increased friction and aerodynamic heating in hypersonic vehicles. This study proposes a novel methodology to control transition in hypersonic wall-bounded flows by introducing spanwise nonuniform temperature surface distributions. The proposed control mechanism modulates the spanwise uniform boundary-layer flow through the generation of spanwise-periodic streamwise-elongated streaks. First, direct numerical simulations are performed at Mach 6 with imposed spanwise nonuniform wall temperature variations. The wall-prescribed temperature conditions successfully generate streaks of sufficient amplitude that suppress transitional instabilities. Second, a wind-tunnel experimental implementation of a steady-state temperature control strategy is proposed, incorporating local cooling below a wall of finite thickness. A thermal model is employed to predict the surface temperature with varying cooling temperature inputs below the finite-thickness wall at Mach 6. The thermal analysis validates the methodology to achieve spanwise nonuniform temperature distributions and emphasizes the sensitivity of temperature input and wall-thickness variations. Lastly, to guide future experimental ground tests, restricted to short test duration, prototype models are proposed in order to regulate the local surface temperature by altering the local heat flux and diffusion within the wall. Specifically, dissimilar materials and thin wall structures are employed to achieve varying levels of heat diffusion and heat flux, respectively. Implementing these approaches generates a higher sensitivity to local surface temperature, thus enabling the passive achievement of spanwise nonuniform temperature distributions, as demonstrated by the thermal model.

Mastering turbulence remains one of physics’ most intractable challenges, with its chaotic, multi-scale dynamics driving energy dissipation across transport and energy systems. Here we report REACT (Reinforcement Learning for Environmental Adaptation and Control of Turbulence), a fully autonomous reinforcement learning framework that achieves real-time, adaptive, closed-loop turbulence control in real-world environments. Deployed on a road vehicle model equipped solely with onboard sensors and servo-actuated surfaces, REACT learns directly from sparse experimental measurements in a wind tunnel environment, bypassing intractable direct numerical simulations and empirical turbulence models. The agent autonomously converges to a policy that reduces aerodynamic drag while achieving net energy savings. Without prior knowledge of flow physics, it discovers that dynamically suppressing spatio-temporally coherent flow structures in the vehicle wake maximizes energy efficiency, achieving two to four times greater performance than model-based baseline controllers. Through a physics-informed training that recasts data in terms of dimensionless physical groups and parametric input spaces, REACT synthesizes offline a single generalizable agent that transfers across speeds without retraining. These results move agentic learning beyond simulation to robust, interpretable real-world control of high‑Reynolds turbulence, opening a path to self‑optimizing physical systems in transport, energy and environmental flows.

The fluid dynamics community has increasingly adopted machine learning to analyze, model, predict, and control a wide range of flows. These methods offer powerful computational capabilities for regression, compression, and optimization. In some cases, machine learning has even outperformed traditional approaches. However, many fluid mechanics problems remain beyond the reach of current machine learning techniques. As the field moves from its current state toward a more mature paradigm, this article offers a critical assessment of the key challenges that must be addressed. Tackling these technical issues will not only deepen our understanding of flow physics but also expand the applicability of machine learning beyond fundamental research. We also highlight the importance of community-maintained datasets and open-source code repositories to accelerate progress in this area. Furthermore, the future success of machine learning in fluid dynamics will depend on effective training—not only for the next generation of researchers but also for established fluid mechanicians adapting to this evolving landscape. Data-driven fluid dynamics is in its critical transitional state over the next few years to shape its future. This perspective article aims to spark discussions and encourage collaborative efforts to advance the integration of machine learning in fluid dynamics.

Mastering turbulence remains one of physics' most intractable challenges, with its chaotic, multi-scale dynamics driving energy dissipation across transport and energy systems. Here we report REACT (Reinforcement Learning for Environmental Adaptation and Control of Turbulence), a fully autonomous reinforcement learning framework that achieves real-time, adaptive, closed-loop turbulence control in real-world environments. Deployed on a road vehicle model equipped solely with onboard sensors and servo-actuated surfaces, REACT learns directly from sparse experimental measurements in a wind tunnel environment, bypassing intractable direct numerical simulations and empirical turbulence models. The agent autonomously converges to a policy that reduces aerodynamic drag while achieving net energy savings. Without prior knowledge of flow physics, it discovers that dynamically suppressing spatio-temporally coherent flow structures in the vehicle wake maximizes energy efficiency, achieving two to four times greater performance than model-based baseline controllers. Through a physics-informed training that recasts data in terms of dimensionless physical groups and parametric input spaces, REACT synthesizes offline a single generalizable agent that transfers across speeds without retraining. These results move agentic learning beyond simulation to robust, interpretable real-world control of high-Reynolds turbulence, opening a path to self-optimizing physical systems in transport, energy and environmental flows.

Optimal transitional mechanisms are analysed for an incompressible shear layer developing over a short, pressure gradient-induced laminar separation bubble (LSB) with peak reversed flow of 2 %. Although the bubble remains globally stable, the shear layer destabilises due to the amplification of external time- and spanwise-periodic disturbances. Using linear resolvent analysis, we demonstrate that the pressure gradient modifies boundary layer receptivity, shifting from Tollmien–Schlichting (T-S) waves and streaks in a zero-pressure-gradient environment to Kelvin–Helmholtz (K-H) and centrifugal instabilities in the presence of the LSB. To characterise the nonlinear evolution of these disturbances, we employ the harmonic-balanced Navier–Stokes (N-S) framework, solving the N-S equations in spectral space with a finite number of Fourier harmonics. Additionally, adjoint optimisation is incorporated to identify forcing disturbances that maximise the mean skin friction drag, conveniently chosen as the cost function for the optimisation problem since it is commonly observed to increase in the transitional stage. Compared with attached boundary layers, this transition scenario exhibits both similarities and differences. While oblique T-S instability is replaced by oblique K-H instability, both induce streamwise rotational forcing through the quadratic nonlinearity of the N-S equations. However, in separated boundary layers, centrifugal instability first generates strong streamwise vortices due to multiple centrifugal resolvent modes, which then develop into streaks via lift-up. Finally, we show that the progressive distortion and disintegration of K-H rollers, driven by streamwise vortices, lead to the breakdown of large coherent structures.