PINN Techniques Interact Nonlinearly, Not Cumulatively, for Fluid Dynamics.

Physics-Informed Neural Networks (PINNs) are gaining traction as an alternative to traditional computational fluid dynamics (CFD) solvers, particularly for complex unsteady flows, by embedding governing equations directly into their training process. Researchers have proposed numerous methods to enhance PINN training, but these are typically validated in isolation, leaving questions about their combined effectiveness. This paper systematically examines how different PINN improvement techniques interact when applied to the DFG/Schafer-Turek unsteady cylinder wake benchmark. Surprisingly, most techniques, when used alone, performed no better than a basic baseline. However, a critical finding emerged: combining periodic (SIREN) activations with causal weighting dramatically improved performance, allowing the reconstruction of velocity and pressure fields with high accuracy (within 4.1% error). Crucially, the study also revealed that simply adding more techniques beyond this effective combination often led to a catastrophic decline in performance. This demonstrates that PINN interventions can interact nonlinearly, and a more elaborate training recipe is not necessarily superior, highlighting the importance of understanding these interactions for effective PINN development.