The oblique detonation wave (ODW) is sustained by the strong coupling between a shock wave and exothermic chemical reactions, resulting in a complex flow structure with sharp moving discontinuities. Although high-fidelity numerical simulations yield detailed physical insights, the associated computational cost remains prohibitive for extensive parametric studies. Classical linear dimensionality reduction methods, such as proper orthogonal decomposition (POD), further suffer from limited capability in representing moving discontinuities due to the slow decay of the Kolmogorov N-width. Motivated by these limitations, this paper proposes a non-intrusive rapid prediction framework based on physics-informed transformation proper orthogonal decomposition coupled with a backpropagation neural network (PiTPOD-BPNN). By extracting physically meaningful feature anchors, the framework aligns the dominant discontinuity structures and maps the non-rigid deformation from the physical space onto a unified reference space. Within this reference space, POD is performed to extract the dominant flow-field modes. A neural network is further employed to learn the nonlinear mapping from the operating parameters to both the POD coefficients and the geometric transformation parameters, thereby enabling rapid flow-field prediction independent of the computational fluid dynamics (CFD) solver. The framework is first validated on a one-dimensional shock-tube problem and a two-dimensional non-reacting oblique shock flow field to assess its capability in capturing moving discontinuities. Subsequently, it is applied to hypersonic ODW flow fields with diverse induction structures. Numerical results demonstrate that the proposed method achieves high-fidelity full-field predictions while substantially reducing computational costs compared to conventional CFD simulations. Notably, for the case involving an abrupt induction structure, the model predicts the locations of the triple point and the ODW exit with relatively small errors. These findings indicate that the PiTPOD-BPNN framework offers an efficient and reliable tool for extensive parametric studies of discontinuity-dominated reactive flows.
The application of Physics-Informed Neural Networks (PINNs) in lubricated contacts, which are typically modeled using elastohydrodynamic lubrication (EHL) simulations, has been investigated and advanced in recent times. This is primarily due to two factors: first, integrating the underlying physics into the network, which accelerates training and reduces the required data; and second, accelerated computations in the investigated test case enabled by rapid evaluations of the trained PINN. Notably, the hydrodynamic (HD) part of an EHL simulation has recently been solved using PINNs by integrating the Reynolds equation into the PINN training. Currently, the research simplifies the HD lubrication problem by assuming stationary conditions, no cavitation, smooth surfaces, a Newtonian fluid, and/or by applying a HD-PINN to a single test case. In this work, an HD-PINN framework is presented that solves the Reynolds equation for rough surfaces, transient cavitation, and non-Newtonian fluids. Furthermore, the framework generates a single parameter-dependent HD-PINN that solves a multi-case problem: several test cases with varying counter-surface velocities for interpolation and extrapolation tasks, thereby accurately expanding the PINN’s solution space beyond the investigated training domain. The results presented underscore the importance and potential of PINNs in lubricated contacts.
A formulation for incompressible magnetohydrodynamic (MHD) flows that eliminates both the pressure Poisson equation and the electric potential Poisson equation is investigated for the simulation of flows under applied magnetic fields. The Entropically Damped Artificial Compressibility (EDAC) method is employed to replace the pressure Poisson equation, while the magnetic induction equation is solved directly in place of the conventional electric potential formulation used for low magnetic Reynolds number flows. The proposed methodology is evaluated using two canonical benchmark problems: the lid-driven cavity (LDC) flow and Rayleigh–Taylor (RT) convection. A comparative convergence study between the EDAC and Poisson-based formulations is performed for the LDC case with an applied magnetic field. The EDAC solver provides significant computational savings while preserving solution fidelity. The influence of axial and transverse magnetic fields on the evolution of RT convection is subsequently examined for varying magnetic interaction parameter, N, and magnetic Reynolds number, Rem. Under an applied transverse magnetic field, increasing N progressively suppresses three-dimensional structures, resulting in a transition towards quasi-two-dimensional flow behavior. In contrast, an applied axial magnetic field suppresses transverse flow structures, producing a characteristic “V”-shaped saddle region accompanied by thinner, finger-like spikes. For both magnetic field configurations, increasing N suppresses vorticity and fluid mixing. The role of Rem is also investigated, with high Rem enhancing the instability growth and fluid mixing, while low Rem suppresses the instability. Overall, the results demonstrate the effectiveness of the EDAC framework for simulating MHD flows and highlight the key roles of magnetic field orientation, N, and Rem in governing buoyancy-driven MHD instabilities.
Granular-ice avalanches in glacierized high-mountain regions can evolve from initially dry, frictional granular-ice mixtures into highly mobile debris flows. This transition is governed by coupled frictional heating, ice melting, pore-pressure generation, and effective-stress reduction. These processes remain insufficiently represented in existing depth-averaged models. In this study, a thermo-mechanically coupled depth-averaged computational framework is developed to simulate the full process of granular-ice avalanches. The moving mass is treated as a three-phase mixture composed of granular particles, ice particles, and pore fluid. Depth-averaged mass and momentum equations are derived for the evolving flow thickness, phase fractions, and mixture velocity. A temperature equation is introduced to describe ice melting. Meltwater production is further coupled with a dilatancy–drainage pore-pressure closure. The framework is evaluated using a sequence of benchmark and validation cases. A one-dimensional dam-break problem verifies the shock-capturing capability, while a granular flow over a simply curved slope confirms the prediction of topography-controlled spreading and frictional arrest. Flume simulations reproduce the measured front velocity and run-out distance of granular-ice mixtures under low to moderate ice-content conditions. A rotating-drum case further demonstrates the progressive transition from initial mechanical adjustment to stable melting and the associated pore-pressure response. The work provides a process-based tool for modelling granular-ice avalanche evolution.
Deep reinforcement learning is employed to discover closed-loop feedback policies for pitching-moment regulation of a NACA0012 airfoil at Rec=3000 using active flow control. Training data are generated from GPU-accelerated simulations of the airfoil with fluidic actuators distributed on its surface. Two actuator configurations are implemented through boundary conditions: wall-normal blowing/suction and Coandă-type wall-tangential blowing. A reward signal drives the agent to achieve aerodynamic trim, defined here as a zero time-averaged pitching moment. The learned feedback policy synchronizes the actuators to periodically generate strategically spaced vortical structures that modify the surface pressure distribution and trim the airfoil. The influence of the time-averaging window in a reward formulation based on a moving average of the pitching moment is assessed. Policies trained with longer averaging windows achieve zero mean pitching moment but do not maintain instantaneous trim. This behavior is mitigated by augmenting the reward with a penalty on the standard deviation of the pitching moment. A post-hoc analysis further shows that the learned policies are effectively linear and considerably simpler than the original neural-network parameterization.
