Aerodynamic analysis in aerospace engineering frequently demands efficient and accurate simulations of complex flows involving moving boundaries. Overset grids offer a flexible approach to grid generation and facilitate the efficient treatment of such boundaries within traditional Computational Fluid Dynamics (CFD). Although Physics-Informed Neural Networks (PINNs) have emerged as promising neural solvers for partial differential equations, their application to overset grid systems remains largely unexplored. Existing methods that extend PINNs to moving boundaries on overset grids typically rely on auxiliary solver data to ensure numerical stability, and their capacity to generalize to unseen flow conditions or geometries has not been rigorously validated. To address these limitations, the first pure data-free physics-informed Graph Neural Network (GNN) framework, Overset-FDGN, is introduced for simulating parametric incompressible flows on overset grids. The two-dimensional unsteady Navier–Stokes equations are discretized via finite differences and embedded directly into the loss function; interactions between subdomains are modeled through message-passing operations. The framework is validated on both stationary cases and moving boundary problems involving forced oscillations: it accurately captures the lock-in phenomenon for a translating cylinder and reliably predicts aerodynamic forces for a pitching airfoil. Furthermore, the model generalizes effectively to previously unseen flow regimes and geometries, achieving close agreement with high-fidelity CFD simulations. Comprehensive evaluations demonstrate that our approach not only matches the accuracy of conventional CFD but also exhibits superior computational efficiency and convergence behavior—while eliminating dependence on external solver data and overcoming the generalization constraints inherent in existing PINN-based overset methods.
Zou et al. (Wed,) studied this question.