Multi-Domain Physics-Informed Machine Learning Based Real-Time Digital-Twin Emulation for a Hydrogen-Powered Maglev Transportation System

Hydrogen-powered maglev trains (HPMLT) are increasingly regarded as promising candidates for next-generation clean rail transportation, owing to their low emissions, potentially reduced maintenance demands, and enhanced operational robustness. However, HPMLT systems exhibit strongly coupled dynamics across the electrical, mechanical, magnetic, and chemical domains, while also involving multi-time-scale behavior arising from the interaction between fast LIM electrical dynamics and comparatively slower magnetic, mechanical, and fuel-cell-related processes. Conventional electromagnetic transient simulations (EMT), predominantly based on numerical integration of detailed physics-based formulations, can become prohibitively time-consuming when confronted with nonlinear multi-domain interactions and fast transients, and may struggle to faithfully preserve cross-domain coupling effects under real-time constraints. To address these limitations, this paper develops a multi-domain, physics-informed real-time digital-twin for an HPMLT system. The proposed approach integrates multi-domain physical constraints grounded in high-fidelity simulation evidence from an Ansys Maxwell finite-element (FE) model and a MATLAB EMT model directly into a physics-informed neural network (PINN) architecture, and subsequently deploys the resulting model on a Xilinx® UltraScale+ VCU118 FPGA platform. Results show that the proposed framework reduces on-chip model-update latency by approximately 57% compared with conventional serial TZR-based approaches, achieving a full HPMLT update in 1.15 s for a 100 s simulation time step. The proposed PINN models maintain mean percentage absolute error below 2% across the considered subsystems, while the PEMFC model exhibits discrepancy below 1% over the investigated operating range. These results demonstrate that the proposed multi-domain PINN-based RTDT provides an effective balance among fidelity, robustness, and deterministic low-latency performance for real-time emulation of strongly coupled HPMLT systems.