Physics-Informed vs. Traditional Neural Networks for Fluid Flow and Heat Transfer

Recent advancements in deep learning have paved the way for solving complex partial differential equations (PDEs) in fluid dynamics. This study presents a comparative analysis between Physics-Informed Neural Networks (PINNs) and Traditional Neural Networks (NNs) for two benchmark problems: Forced Convection in an Enclosure and Flow Past a Cylinder. These problems were selected because they are canonical cases in computational fluid dynamics, widely used to evaluate numerical methods due to their well-understood physical behavior and relevance to practical applications. The primary objective is to assess the performance of both approaches in terms of accuracy, generalization, and physical consistency when pre- dicting key fluid dynamics variables such as velocity fields, pressure distributions, and temperature profiles. In the Forced Convection problem, PINNs demonstrated superior performance by accurately capturing the temperature field and streamlines, enforcing physical laws directly through the governing equations. Traditional NNs showed limitations in learning the underlying physics, resulting in asymmetric and less realistic temperature distributions. In the Flow Past a Cylinder problem, PINNs effectively modeled the pressure field and streamlines with better physical consistency, capturing the pressure gradient and flow separation. In contrast, NNs produced periodic and unrealistic flow patterns, failing to capture critical features such as vortex shedding and recirculation zones. Loss curves indicate faster convergence and better generalization of PINNs compared to NNs, leveraging embedded physics knowledge through the Navier–Stokes and heat transfer equations