Using Deep Learning for Predictive Modelling in Fluid Dynamics

Predictive modeling of fluid dynamics is vital for many engineering applications, but classical computational fluid dynamics (CFD) methods tend to suffer with high computing costs and low scalability, notably in the complicated turbulent flow regimes. In this work we investigate the use of deep learning models-specifically Convolutional Neural Networks (CNNs), Graph Neural Networks (GNNs), and Physics-Informed Neural Networks (PINN s)-to develop predictive models that may supplement fluid dynamics simulations with potentially im-proved accuracy and efficiency. Perform an evaluation and comparison of the forecasting ability for the velocity fields of these models with an emphasis on turbulent flow regimes. The authors apply each model to a real-world fluid dynamics dataset, assessing their success across multiple axes: large-scale feature capture, localized interactions, and physical law adherence. However, the maj or findings demonstrate that PINN s had the best performance with a coefficient of determination (R2) of 0.97 and mean absolute error (MAE) of 0.015 across turbulent regions but relatively weaker CNNs and GNNs with an accuracy figure of 95%. Although CNNs were successful in modeling the large-scale flow features and GNN s were easy to extract localized interactions, PINN s delivered the most accurate results by adding the N avier-Stokes equations into the loss functions [11]. Abstract Smooth and accurate solutions are vital in fluid dynamics, especial-Iy for achieving physical consistency. This work highlights the potential of PINNs for obtaining such predictions, giving a pathway towards efficient simulations in practical applications.