Physics-informed machine learning modeling and inferencer-in-the-loop based real-time digital-twin emulation for a Maglev transportation system

In response to the increasing demands for efficient, sustainable, and high-speed transportation systems, magnetic levitation (Maglev) train technology has garnered significant attention due to its frictionless operation, high efficiency, low maintenance and environmental benefits. However, accurately capturing the complex interactions among electrical, mechanical, and disparate electromagnetic components remains a significant modeling challenge. This paper introduces a physics-informed machine learning (PIML)-based modeling framework specifically designed for Maglev systems, enhanced by real-time digital-twin emulation implemented on a Xilinx UltraScale+ VCU118 FPGA platform. The proposed physics-informed partitioned neural network (PIPNN) leverages physical principles within its architecture, providing superior predictive accuracy, interpretability, and computational efficiency compared to conventional electromagnetic transient (EMT) methods and traditional neural network models. Comprehensive emulations validate the PIPNN's capability to emulate Maglev system dynamics in real-time with a prediction error consistently below 5%, highlighting errors as low as 2% for specific subsystem variables. Additionally, the hybrid offline-online training strategy and the real-time inferencer-in-the-loop (IIL) digital-twin framework enhance the model's adaptability and robustness under varying operational conditions, demonstrating significant potential for practical deployment in advanced transportation systems.