Predicting ride comfort in high-speed rail using machine learning

Ride comfort is a necessary criterion for assessing the passenger experience in next-generation high-speed rail (HSR), especially as speeds exceed 300 km/h. However, accurate comfort prediction under varying track and speed conditions remains a computational challenge when relying solely on physics-based simulation. This study proposes a surrogate ensemble meta-model trained on vibration responses generated by a validated multi-body vehicle–track interaction simulation. It uses speed, track irregularity type, and wavelength as key inputs and predicts frequency-weighted root-mean-square (RMS) acceleration in accordance with ISO 2631-1. Five ensemble learning models were compared and fused into a stacked Random Forest meta-model, which achieved a predictive R² of 92.1% and a mean absolute error of 0.013 m/s² on unseen cases. Learning curves and prediction intervals confirm strong generalization despite the modest dataset size (n = 345), with the model capturing nonlinear comfort deterioration between 300–400 km/h and identifying the critical 50–100 m wavelength band. This simulation-aligned machine learning surrogate provides a computationally efficient alternative to repeated multi-body simulations, enabling rapid evaluation of speed–track interaction scenarios within the validated modeling framework. Rather than replacing physics-based analysis, the proposed approach is intended to support early-stage design screening, sensitivity analysis, and large-scale scenario exploration where full simulation would be computationally prohibitive. These results establish the feasibility of integrating data-driven surrogates with validated vehicle–track models to accelerate comfort-oriented engineering assessments in HSR systems.