Physics-Informed Neural Network Framework for Wheel-Rail Contact Analysis: Toward an Intelligent Maintenance Workflow in Railway Systems

Railway transportation traffic is rapidly growing, which demands a more effective and efficient generation analysis for more reliable predictive maintenance planning. This is achievable if prognostic indicators are known, e.g. stress, deformation and displacement fields. These parameter values unlock the underlying physics knowledge about fault modes and mechanisms to solve the wheel-rail deterioration. Commercial multi-physics software lack source code accessibility, flexibility and interoperability between computing platforms. On the other hand, physics-informed neural networks (PINNs), which belong to the second AI revolution and scientific ML (SciML) that combines physical and machine learning models, show promise in computational fluid dynamics and electrodynamics; however, their application to railwayremains largely unexplored. This study addresses these research gaps through a comprehensive open-source and reproducible PINN PhysicsNeMo framework for 3D wheel-rail contact analysis as proof of concept. Current railway maintenance often relies on reactive approaches; this PhysicsNeMo framework supports integration by providing predictive stress analyses. The aim is to explore the PhysicsNeMo simulations for railway, establishing a foundation for an interpretable, explainable and trustworthy AI. Results demonstrate detailed and intuitive 3D wheel visualisations of stress distributions and displacement fields, with insights into damage mechanisms for railway designers and maintainers, facilitating more efficient maintenance workflows.