Speaker
Description
Accurate simulation of unsteady fluid dynamics is critical for applications ranging from aerospace engineering to climate modelling. However, the prohibitive computational cost of high-fidelity solvers often precludes their use in real-time control. Traditionally, Finite Volume Method (FVM) solvers, such as Open FOAM, haveserved as the gold standard for accuracy but they are highly dependent on dense mesh discretization’s which creates a computational bottleneck. Reduced Order Models (ROMs) and Physics Informed Neural Networks (PINNs) have emerged as promising alternatives which acts as high-speed surrogates. In this work, we present a comparative analysis of FVM, POD-Galerkin ROMs, and PINNs applied to the canonical benchmark of unsteady flow around a cylinder at Re = 100. We demonstrate that the POD-Galer kin ROM technique reduces computational time forty-fold, slashing simulation du ration from 48 minutes to just 73 seconds, yet its utility remains strictly confined to the parametric range of the training snapshots. Conversely, we find that while PINNs offer mesh-agnostic flexibility, they often struggle to capture high-frequency vortex shedding due to spectral bias. Crucially, we show that augmenting PINNs with the ADAM-LFBGS optimizer effectively mitigates these stability issues, allowing them to match the fidelity of projection-based methods.