Machine learning for safety distances prediction during emergency response of toxic dispersion accidental scenarios

Abstract

The present work investigates the prediction of safety distances (SDs) for toxic dispersion scenarios using machine learning (ML) models and examines an application to emergency response cases. PHAST software was used to generate multiple dispersion scenarios over a wide range of ambient conditions, leak and catastrophic rupture loss of containment focusing on nine toxic substances in varying storage conditions. The database underwent exploratory analysis and dependency testing to thoroughly analyze its content and the interdependencies among the input features. A correlation matrix facilitated feature reduction by identifying redundant features which were disregarded. The nine input features used were molecular weight (Mw), mass, storage phase, storage temperature, air temperature, stability class, wind, ground roughness and hole size. Following, four ML models, namely Random Forest, AdaBoost, HistGradientBoosting and LightGBM, were trained and optimized on the database. Their performance was compared across three sets (validation, test and extrapolation) using two common statistical metrics (RMSE and R²) and a custom metric (CS) designed to address the specific characteristics of SDs. Among the four models, LightGBM showed the best performance with an RMSE of 20.00 m and a CS of 3.65 m in the test set. Sensitivity analysis techniques (i.e., permutation importance, Sobol S1 and DMIM indices) were used to rank the importance of the features of the LightGBM model. This model was then integrated into a logic diagram capable of handling cases where some input features are unknown. Seven hypothetical cases were defined to demonstrate the practicability of such tool for emergency response.