Hypersonic Flow Control: Generalized Deep Reinforcement Learning for Hypersonic Intake Unstart Control under Uncertainty

TL;DR

This work tackles hypersonic inlet unstart at $M_ty=5$, where rapid back-pressure rise can destabilize shock trains and trigger unstart. It combines a high-fidelity fifth-order spectral DG CFD solver with adaptive mesh refinement and off-policy DRL (SAC/TD3) to learn closed-loop microjet actuation policies that stabilize the shock system across multiple throttling ratios and sensor configurations. The DRL controller demonstrates strong zero-shot generalization to unseen back-pressures and Reynolds numbers, maintains performance under 5–10% sensor noise, and identifies an optimal sparse sensor set for practical implementation, achieving real-time operation at $50$ kHz with $20$ μs update intervals. These results indicate a viable data-driven pathway toward robust, real-time hypersonic intake control under realistic uncertainties, paving the way for integration with physics-aware surrogates and real-time hardware accelerators for flight testing.

Abstract

The hypersonic unstart phenomenon poses a major challenge to reliable air-breathing propulsion at Mach 5 and above, where strong shock-boundary-layer interactions and rapid pressure fluctuations can destabilize inlet operation. Here, we demonstrate a deep reinforcement learning (DRL)- based active flow control strategy to control unstart in a canonical two-dimensional hypersonic inlet at Mach 5 and Reynolds number $5\times 10^6$. The in-house CFD solver enables high-fidelity simulations with adaptive mesh refinement, resolving key flow features, including shock motion, boundary-layer dynamics, and flow separation, that are essential for learning physically consistent control policies suitable for real-time deployment. The DRL controller robustly stabilizes the inlet over a wide range of back pressures representative of varying combustion chamber conditions. It further generalizes to previously unseen scenarios, including different back-pressure levels, Reynolds numbers, and sensor configurations, while operating with noisy measurements, thereby demonstrating strong zero-shot generalization. Control remains robust in the presence of noisy sensor measurements, and a minimal, optimally selected sensor set achieves comparable performance, enabling practical implementation. These results establish a data-driven approach for real-time hypersonic flow control under realistic operational uncertainties.

Figures (23)

• Figure 1: Computational domain $(\Omega)$ with inflow Mach 5 and boundary conditions. All dimensions are in 'mm'.
• Figure 2: TR34: Mach number contours at successive time steps illustrating the gradual buildup of back pressure, followed by subsonic spillage and the onset of inlet unstart at Mach 5 and $R_e = 5 \times 10^6$. From top to bottom, the rows correspond to times $t = 0.1, 0.5, 1.0, 2.0, 3.0, 3.3, 3.4,$ and $3.8~\mathrm{ms}$. The first column shows the Mach number contours, while the second column shows the corresponding AMR distribution. Points 1 and 2 (top-right) shows the location where temporal evolution of pressure is measured.
• Figure 3: TR34: hp-convergence study of the pressure ratios $p_1/p_{\infty}$ and $p_2/p_{\infty}$ for polynomial orders $5$, $6$, and $7$. The results demonstrate convergence of the solution with increasing polynomial order as well as through mesh refinement using AMR. The approximate number of elements for 5th, 6th, and 7th polynomial orders is 40k, 50k, and 60k, respectively.
• Figure 4: Effect of throttling on the pressure ratios $p_1/p_{\infty}$ and $p_2/p_{\infty}$ for TR30, TR34, TR38, and TR40, including the unthrottled condition. The variation of $p_2/p_{\infty}$ clearly indicates the onset of the upstart phenomenon, characterized by a sudden rise in pressure. Furthermore, the onset of upstart occurs earlier in time as the TR increases, which is attributed to the corresponding increase in back pressure.
• Figure 5: Schematic of microjet placement along the wall: blowing microjets with learnable jet angle ($\beta$) are applied on $\Gamma_1$, whereas suction microjets are applied on $\Gamma_2$ and $\Gamma_3$. Pressure sensor points are indicated by blue rectangles located within the isolator.