PHYSICS-INFORMED NEURAL ARCHITECTURES FOR REAL PREDICTION OF TURBULENT FLOW IN HYPERSONIC BOUNDARY LAYERS

Abstract

Research AbstractMotivation and Purpose of ResearchThe prediction of turbulent boundary layers in hypersonic flight regimes (Ma \ge 5) is critically hindered by the computational constraints of classical fluid dynamics. While Direct Numerical Simulation (DNS) provides the high-fidelity data required to understand extreme aerodynamic heating and drag amplification, it demands millions of CPU-core hours, rendering it unviable for real-time in-flight control. Conversely, faster Reynolds-Averaged Navier-Stokes (RANS) models consistently fail at hypersonic speeds due to their inability to resolve non-linear turbulence-chemistry interactions and compressibility effects. The primary objective of this research is to bridge this gap by developing a Physics-Informed Neural Network (PINN) capable of predicting hypersonic aerothermal loads and turbulent flow behaviour in real time, bypassing the grid-scaling limitations of traditional solvers entirely.Research Methods and ResultsThe proposed framework utilises a deep neural network trained on sparse, high-fidelity DNS datasets generated via the CharLES solver. Instead of relying solely on data-driven curve fitting, the PINN embeds the full compressible Navier-Stokes equations directly into its composite loss function. This enforces the strict conservation of mass, momentum, and energy via automatic differentiation. The network's hyperparameters were refined through Bayesian Optimisation, and gradient stiffness was mitigated using Neural Tangent Kernel (NTK) balancing.The results demonstrate that the PINN accurately predicts velocity, temperature, and pressure fields across Mach 5, 7, and 10 regimes with a mean absolute percentage error (MAPE) of less than 3%. Crucially, the inference latency was reduced from approximately 48 hours (DNS) to under 50 milliseconds on a single GPU—a 3.5 \times 10^6 magnitude speedup. Furthermore, via transfer learning, the model successfully generalised to complex geometries, such as a 25-degree compression corner, pinpointing shock-wave reattachment points to within 1.3% of the DNS baseline.Conclusions and Future PlanThis research confirms that PINNs can effectively map the continuous solution manifold of hypersonic turbulence, delivering the physical accuracy of DNS at a fraction of the computational cost. This architectural leap enables the real-time calculation of skin-friction coefficients and heat fluxes necessary for onboard closed-loop control. Future plans involve integrating this PINN framework into Model Predictive Control (MPC) loops for active regenerative cooling in scramjet engines. Additionally, the methodology will be expanded to cross-domain applications, specifically targeting magnetohydrodynamic (MHD) plasma turbulence containment in nuclear fusion reactors.

Keywords

Natural Science - Physics

Citation

References

Reference metadata is pending and must be finalized before DOI deposit.