PINN-Based CFD Prediction Compared with ANSYS Fluent

A student-laptop study of how well physics-informed neural networks (PINNs) reproduce steady laminar CFD solutions. Three benchmarks - Kovasznay flow at Re 20 (exact solution), the lid-driven cavity at Re 100 (Ghia et al. 1982) and the channel flow past a cylinder at Re 20 (Schafer-Turek 2D-1) - are solved by an own finite-volume solver (the stand-in for ANSYS Fluent; cylinder Cd 5.526 against 5.5795) and by 6 x 64 tanh PINNs trained on the Navier-Stokes residuals with Adam and L-BFGS on a shared NVIDIA GeForce RTX 2050. The physics-only PINN matches the Kovasznay field within 0.07 % (u), reaches 10.0 % speed error in the cavity and 2.3 % when 64 CFD velocity sensors are added, while a data-only network without physics fails on the pressure. A parametric PINN predicts the cavity for any Re between 50 and 400. A role-based web dashboard shows every field, profile and coefficient. Honest scope: ANSYS is NOT installed on the build PC - the Fluent journals, Fluent Meshing journal, SpaceClaim scripts and UDFs were written but never run; a real Fluent export dropped into data/fluent/<case>.csv replaces the reference automatically.

Category: Mechanical Engineering / Computational Fluid Dynamics / Scientific Machine Learning / Physics-Informed Neural Networks / Surrogate Modelling / Verification and Validation / ANSYS Fluent (journals) with Python and PyTorch

Mechanical Engineering Project M04 from ProTechSC Academy, with graphical abstract, video demonstration and project guidance.