Explainable Machine Learning for Predicting Time‐to‐Failure of Horizontal Pressurized Storage Tanks in Fire Scenarios

ABSTRACT Fire‐induced failure of pressurized hazardous chemical storage tanks under elevated temperature and pressure may trigger severe domino accidents, making accurate time‐to‐failure (TTF) prediction essential for preventing accident escalation and assessing tank vulnerability. However, interpretable machine learning (ML) models for predicting the TTF of horizontal pressurized storage tanks remain underexplored. This study proposes a data‐driven framework for predicting the TTF of horizontal pressurized tanks exposed to pool fires. Using typical tank geometries specified in Chinese standards and representative fire parameters from the literature, a comprehensive TTF dataset was generated via a lumped‐parameter method and used to train and validate three ML models: fully connected feedforward neural network (FNN), random forest (RF), and XGBoost. Results indicate that the FNN significantly outperforms the other two ML models and a probit‐based simplified correlation in terms of accuracy, robustness, and normality of error distribution. SHapley Additive exPlanations (SHAP) is further introduced to quantify and rank feature contributions to both model output and predictive uncertainty, with dependence analysis additionally revealing nonlinear behaviors and threshold effects of key features. In addition, a simplified explicit correlation was fitted to the data generated by RADMOD, enabling rapid estimation of TTF in industrial applications where a full neural network cannot be deployed. The proposed framework provides an accurate and physically interpretable tool for the quantitative risk assessment of fire‐induced failure in horizontal pressurized storage tanks.