Machine Learning-Based Geometric Interpolation for Fluid-Flow Modelling

Modelling fluid flows accurately is essential in engineering and science, but it often depends on solving complex Partial Differential Equations (PDEs), which can be computationally demanding.In this study, we leverage Artificial Neural Networks (ANNs) as efficient surrogates for traditional numerical methods, predicting fluid-flow characteristics with significantly reduced computational cost.Building on our recent previous work, we extend our analysis to explore geometric interpolation within a two-dimensional (2D) channel.While we investigate interpolation with respect to the geometry of the fluid-flow problem, we explore extrapolation in terms of Reynolds numbers, thereby validating our Machine-Learning model on fluid flows with Reynolds numbers beyond the training range.The obtained results demonstrate that while training on finer grids yields high accuracy, training on coarser grids still achieves comparable performance with minimal computational effort; this outcome highlights in turn the promise of ANNs for efficient modelling of fluid dynamics.Such capabilities are particularly valuable in contemporary applications such as digital twins, enabling real-time system monitoring and optimization.