TransPhyX: A data-driven method for dynamic physical field prediction in stochastic load time-series

Abstract

Dynamic prediction of in-service physical fields (e.g. stress, strain, and temperature fields) constitutes a cornerstone technology for digital governance of mechanical equipment. The stochasticity and time-varying characteristics of external excitation loads (e.g. thermal, vibrational, and impact loads) introduce significant complexity in physical field prediction. Online monitoring of physical fields at the assembly interfaces of mechanical systems is critical for ensuring structural safety, extending service life, and optimizing design. This study proposes TransPhyX (Transformer-Based Physical Field Prediction with XGBoost Precoder), a hybrid data-driven framework designed to overcome these challenges. The novelty of TransPhyX lies in: (1) a recursive stochastic load generation and parametric dataset construction method tailored for dynamic prediction tasks; (2) a modular hybrid architecture that decouples transient load encoding (via XGBoost) and dynamic sequence modeling (via Transformer), improving spatiotemporal continuity and generalization; and (3) an Outlier Removal Ensemble (ORE) algorithm that fuses multi-scale predictions to eliminate anomalies and enhance robustness. Validated on flip-chip thermal management and flange-bolt stress prediction, TransPhyX achieves 99.79 % prediction fidelity with a 97.79 % reduction in computational costs compared to FEM, outperforming AutoGAN and TransUNet baselines in both accuracy and stability. These contributions establish TransPhyX as a rapid, high-fidelity solution for real-time structural health monitoring and digital twin implementation in stochastic loading environments.