Physics-consistent alignment and discrepancy learning for multi-fidelity prediction of train collision load

Train collision accidents may lead to severe casualties and substantial property losses, making crashworthiness analysis essential for improving train passive safety. However, the high computational cost of high-fidelity numerical simulations limits their direct application to large-scale parameter analysis and optimization. To address this challenge, this study proposes a bias-correction surrogate network (BCS-Net) for multi-fidelity prediction of train collision loads under limited high-fidelity data. The proposed framework performs a physically consistent alignment of force-displacement responses from models with different fidelity levels within a unified displacement domain. On this basis, systematic discrepancies between low- and high-fidelity models are captured through low-dimensional modeling and learning, enabling effective multi-fidelity information fusion. Furthermore, the bias-corrected surrogate is embedded into a multi-objective crashworthiness optimization framework to support efficient parameter configuration. The results demonstrate that the proposed BCS-Net significantly improves force-displacement curve prediction accuracy while maintaining a substantial computational efficiency advantage over high-fidelity simulations. Its effectiveness is further validated through multi-objective crashworthiness parameter optimization. The proposed framework provides a practical and reliable surrogate-based solution for train crashworthiness analysis and parameter optimization.