Accelerated physics-based simulations of train aerodynamics using machine learning libraries

• Neural Physics, a CNN-based approach with weights from classical discretisation, enables high-fidelity train aerodynamic simulations on GPUs and AI processors with high speed and accuracy. • The method integrates implicit large eddy simulation with a non-linear Petrov-Galerkin scheme and uses ConvFEM for higher-order finite elements, simplifying implementation. • First application of ConvFEM to high-speed fluid flow in complex geometries, demonstrated on a freight Locomotive Class 66 and a partially loaded freight train with and without crosswind. • Validated results show good agreement with existing numerical and experimental data for pressure and velocity distributions, indicating potential for efficient predictive frameworks in train-environment interaction studies.