Dynamic Erosion Prediction of Tundish Refractories Based on Hybrid Deep Learning

To quantify the CFD‐resolved thermal–mechanical erosion of tundish refractories in continuous steel casting, this study proposes a digital twin early warning model integrating multiphysical fields and hybrid deep learning. Combining CFD simulations and on‐site data, physical features were extracted and screened using PCC‐LASSO, with unreliable pressure boundaries actively removed. A PSO‐optimized BiLSTM–Transformer model was then constructed. By introducing target lag features and smoothed thermodynamic features state, the model synergistically captures both local temporal fluctuations and global cumulative erosion effects, overcoming the extrapolation limits of traditional regression models. Additionally, Monte Carlo Dropout is utilized to dynamically output a 95% prediction confidence interval for quantified risk assessment. Experiments demonstrate that the proposed model achieves a significantly lower maximum absolute error (MaxAE) than benchmarks like SVR and XGBoost, exhibiting strong generalization across unseen operating conditions. Furthermore, SHAP attribution analysis shows a feature‐attribution pattern consistent with turbulence‐enhanced convective heat transfer driving the CFD‐resolved thermal–mechanical thinning within the enthalpy–porosity description. Ultimately, this research provides a highly reliable algorithmic pathway for predictive maintenance in complex metallurgical scenarios.