CFD-based recurrent neural network modeling of plasma-assisted flow control over a stepped NACA4412 airfoil using LSTM and GRU

Abstract This study explores a hybrid computational data driven framework for analyzing and predicting unsteady aerodynamic flow behavior over a NACA4412 airfoil incorporating backward-facing step (BFS) modifications and dual alternating current dielectric barrier discharge (AC-DBD) plasma actuators. High fidelity numerical simulations were conducted at a Reynolds number of $$3.1 \times 10^{6}$$ for angle of attack (AOA) ranging from $$0^\circ$$ to $$20^\circ$$ using the SST k – $$\omega$$ turbulence model to generate time resolved flow field data. BFS configurations located on the pressure side near the trailing edge with step lengths of $$10\%c$$ , $$20\%c$$ and $$30\%c$$ , were systematically examined under both baseline and plasma actuated conditions. Aerodynamic performance and flow dynamics were quantified through lift and drag coefficients, lift-to-drag ratio and vorticity magnitude to assess separation characteristics and turbulence evolution. The results reveal that while standalone BFS geometries promote lift enhancement by energizing the near wall flow and they also incur a moderate increase in aerodynamic drag. The introduction of dual AC-DBD plasma actuation markedly improves flow control performance by delaying boundary-layer separation, weakening large-scale vortical structures and stabilizing the shear layer. The combined configuration achieved a maximum lift improvement of $$83.07\%$$ for the $$30\%c$$ BFS case and a peak lift-to-drag ratio enhancement of $$44.43\%$$ for the $$10\%c$$ BFS case. To facilitate rapid flow field prediction the long short-term memory (LSTM) and gated recurrent unit (GRU) neural networks were trained on computational fluid dynamics generated velocity data. The trained models accurately predicted future velocity contours, demonstrating strong agreement with numerical simulations and effectively capturing dominant unsteady flow features. These findings highlight the capability of machine learning based surrogate models to efficiently reproduce complex aerodynamic responses associated with hybrid flow control strategies, offering a promising pathway for accelerated aerodynamic analysis and control-oriented applications.